fio(1)

📦

version 3.42 · 9 versions

3.42 · arch rolling, nixpkgs unstable (latest)
3.38 · opensuse-leap 16.0
3.35 · centos-stream 9
3.28 · ubuntu jammy
3.40 · fedora 43, fedora 44
3.33 · debian bookworm
3.36 · centos-stream 10, ubuntu noble
3.41 · alpine 3.24, nixos 26.05, ubuntu resolute, ubuntu stonking
3.39 · debian trixie, ubuntu questing

💅 fio(1) — the general commands manual, but make it make sense 🎀

okay so bestie, we're talking about fio, from the General Commands Manual section, and honestly? she's iconic. let's get into it.

✨ NAME ✨

fio — flexible I/O tester. that's the whole vibe. flexible. I/O. tester. periodt. 💖

🌸 SYNOPSIS 🌸

fio [options] [jobfile]...

🦋 DESCRIPTION 🦋

so fio is a tool that spawns a bunch of threads or processes doing a particular type of I/O action — exactly the kind you specify. think of it as a little I/O army that does your bidding. ✨

the typical way to use fio is to write a job file that matches the I/O load you want to simulate. that's the main character move, fr fr.

👑 OPTIONS 👑

all the flags, no gatekeeping. here we go:

--debug=type — enables verbose tracing of type for various fio actions. you can pass all for every type, or list individual types separated by a comma (so like --debug=file,mem enables file and memory debugging, slay). help will list all the available tracing options. 💅
--parse-only — parse options only, don't start any I/O. reading the menu, not ordering. 🙈
--merge-blktrace-only — merge blktraces only, don't start any I/O.
--output=filename — write output to filename.
--output-format=format — set the reporting format to normal, terse, json, or json+. you can select multiple formats, just separate them with a comma. terse is a CSV based format. json+ is like json, except it adds a full dump of the latency buckets. fancy.
--bandwidth-log — generate aggregate bandwidth logs. filename is an optional argument. if you don't give it, fio will by default create files called "agg-{read,write,trim}_bw.log". if you DO give the filename argument, fio creates files called "{filename}-{read,write,trim}_bw.log". get you a log that can do both. 💖
--minimal — print statistics in a terse, semicolon-delimited format.
--append-terse — print statistics in the selected mode AND terse, semicolon-delimited format. ⚠️ deprecated, bestie — use --output-format instead to select multiple formats.
--terse-version=version — set the terse version output format (default 3, or 2, 4, 5).
--version — print version information and exit.
--help — print a summary of the command line options and exit. the classic.
--cpuclock-test — perform test and validation of the internal CPU clock.
--crctest=[test] — test the speed of the built-in checksumming functions. if no argument is given, all of them are tested. alternatively you can pass a comma separated list, in which case the given ones are tested.
--cmdhelp=command — print help information for command. may be all for all commands.
--enghelp=[ioengine[,command]] — list all commands defined by ioengine, or print help for command defined by ioengine. if no ioengine is given, list all available ioengines.
--showcmd — convert given jobfiles to a set of command-line options.
--readonly — turn on safety read-only checks, preventing writes and trims. this is the extra safety guard that stops users from accidentally starting a write or trim workload when that's not what they wanted. fio will only modify the device under test if rw=write/randwrite/rw/randrw/trim/randtrim/trimwrite is given. use this safety net as an extra precaution, bestie. 🫶
--eta=when — specifies when the real-time ETA estimate should be printed. when may be always, never or auto. auto is the default — it prints ETA when requested if the output is a TTY. always disregards the output type and prints ETA when requested. never never prints ETA. straightforward queen behavior.
--eta-interval=time — by default, fio requests client ETA status roughly every second. with this option, the interval is configurable. fio imposes a minimum allowed time to avoid flooding the console — less than 250 msec is not supported. ⏱️
--eta-newline=time — force a new line for every time period passed. when the unit is omitted, the value is interpreted in seconds.
--status-interval=time — force a full status dump of cumulative (from job start) values at time intervals. this option does not provide per-period measurements, so values like bandwidth are running averages. when the time unit is omitted, time is interpreted in seconds. ⚠️ note that using this with --output-format=json will yield output that technically isn't valid json, since the output will be collated sets of valid json — it'll need to be split into valid sets of json after the run.
--section=name — only run the specified section name in the job file. multiple sections can be specified. this option lets you combine related jobs into one file — like, one job file could define light, moderate, and heavy sections, and you tell fio to run only the "heavy" section by giving --section=heavy on the command line. you can also specify the "write" operations in one section and the "verify" operation in another. the --section option only applies to job sections — the reserved global section is always parsed and used. 👑
--alloc-size=kb — allocate additional internal smalloc pools of size kb in KiB. this option increases shared memory set aside for use by fio. if you're running large jobs with randommap enabled, fio can run out of memory. smalloc is an internal allocator for shared structures from a fixed size memory pool and can grow to 16 pools. the pool size defaults to 16MiB. NOTE: while running, .fio_smalloc.* backing store files are visible in /tmp. 👀
--warnings-fatal — all fio parser warnings are fatal, causing fio to exit with an error. no soft launches here. 💀
--max-jobs=nr — set the maximum number of threads/processes to support to nr. NOTE: on Linux, it may be necessary to increase the shared-memory limit (/proc/sys/kernel/shmmax) if fio runs into errors while creating jobs.
--server=args — start a backend server, with args specifying what to listen to. see the CLIENT/SERVER section.
--daemonize=pidfile — background a fio server, writing the pid to the given pidfile file.
--client=hostname — instead of running the jobs locally, send and run them on the given hostname or set of hostnames. see the CLIENT/SERVER section.
--remote-config=file — tell the fio server to load this local file.
--idle-prof=option — report CPU idleness. option is one of the following:
- calibrate — run unit work calibration only and exit.
- system — show aggregate system idleness and unit work.
- percpu — as system but also show per CPU idleness.
--inflate-log=log — inflate and output compressed log.
--trigger-file=file — execute trigger command when file exists.
--trigger-timeout=time — execute trigger at this time.
--trigger=command — set this command as local trigger.
--trigger-remote=command — set this command as remote trigger.
--aux-path=path — use the directory specified by path for generated state files instead of the current working directory.

🎀 JOB FILE FORMAT 🎀

okay so any parameters following the options will be assumed to be job files — unless they match a job file parameter. multiple job files can be listed, and each job file will be regarded as a separate group. fio will stonewall execution between each group. the drama. 💅

fio accepts one or more job files describing what it's supposed to do. the job file format is the classic ini file, where the names enclosed in [] brackets define the job name. you're free to use any ASCII name you want — except global, which has special meaning. following the job name comes a sequence of zero or more parameters, one per line, that define the behavior of the job. if the first character in a line is a ; or a #, the entire line is discarded as a comment. 🙈

a global section sets defaults for the jobs described in that file. a job may override a global section parameter, and a job file may even have several global sections if so desired. a job is only affected by a global section residing above it. iykyk — position matters. ✨

the --cmdhelp option also lists all options. if used with a command argument, --cmdhelp will detail the given command.

see the examples/ directory for inspiration on how to write job files. note the copyright and license requirements currently apply to examples/ files. 📝

also, the maximum length of a line in the job file is 8192 bytes. don't be out here writing a novel on one line, babes.

💜 JOB FILE PARAMETERS 💜

some parameters take an option of a given type, like an integer or a string. anywhere a numeric value is required, an arithmetic expression may be used — provided it's surrounded by parentheses. supported operators are:

addition (+)
subtraction (-)
multiplication (*)
division (/)
modulus (%)
exponentiation (^)

for time values in expressions, units are microseconds by default. 🫠 this is different than for time values not in expressions (not enclosed in parentheses). keep that in mind, bestie. 🦋

💖 PARAMETER TYPES 💖

okay so these are the parameter types fio uses, bestie — think of them like the little labels that tell you what flavor of value each thing wants ✨

str — String. A sequence of alphanumeric characters.
time — Integer with a possible time suffix. If you don’t give a unit, the value is interpreted as seconds unless otherwise specified. It accepts a suffix of d for days, h for hours, m for minutes, s for seconds, ms (or msec) for milliseconds, and us (or usec) for microseconds. For example, use 10m for 10 minutes.
int — Integer. A whole number value, which may contain an integer prefix and an integer suffix.

[*integer prefix*] **number** [*integer suffix*]

The optional integer prefix specifies the number’s base. The default is decimal. 0x specifies hexadecimal.

The optional integer suffix specifies the number’s units, and includes an optional unit prefix and an optional unit. For quantities of data, the default unit is bytes. For quantities of time, the default unit is seconds unless otherwise specified.

With kb_base=1000, fio follows international standards for unit prefixes. To specify power-of-10 decimal values defined in the International System of Units (SI):

K means kilo (K) or 1000
M means mega (M) or 1000**2
G means giga (G) or 1000**3
T means tera (T) or 1000**4
P means peta (P) or 1000**5

To specify power-of-2 binary values defined in IEC 80000-13:

Ki means kibi (Ki) or 1024
Mi means mebi (Mi) or 1024**2
Gi means gibi (Gi) or 1024**3
Ti means tebi (Ti) or 1024**4
Pi means pebi (Pi) or 1024**5

For Zone Block Device Mode:

z means Zone

With kb_base=1024 (the default), the unit prefixes are opposite from those specified in the SI and IEC 80000-13 standards to provide compatibility with old scripts. For example, 4k means 4096.

For quantities of data, an optional unit of B may be included (e.g., kB is the same as k).

The integer suffix is not case sensitive (e.g., m/mi mean mebi/mega, not milli). b and B both mean byte, not bit.

Examples with kb_base=1000:

4 KiB: 4096, 4096b, 4096B, 4k, 4kb, 4kB, 4K, 4KB
1 MiB: 1048576, 1m, 1024k
1 MB: 1000000, 1mi, 1000ki
1 TiB: 1073741824, 1t, 1024m, 1048576k
1 TB: 1000000000, 1ti, 1000mi, 1000000ki

Examples with kb_base=1024 (default):

4 KiB: 4096, 4096b, 4096B, 4k, 4kb, 4kB, 4K, 4KB
1 MiB: 1048576, 1m, 1024k
1 MB: 1000000, 1mi, 1000ki
1 TiB: 1073741824, 1t, 1024m, 1048576k
1 TB: 1000000000, 1ti, 1000mi, 1000000ki

To specify times (units are not case sensitive):

D means days
H means hours
M means minutes
s or sec means seconds (default)
ms or msec means milliseconds
us or usec means microseconds

The z suffix specifies that the value is measured in zones. Value is recalculated once the block device’s zone size becomes known.

If the option accepts an upper and lower range, use a colon : or minus - to separate such values. See the irange parameter type. If the lower value specified happens to be larger than the upper value, the two values are swapped.

bool — Boolean. Usually parsed as an integer, however only defined for true and false (1 and 0).
irange — Integer range with suffix. Allows a value range to be given, such as 1024-4096. A colon may also be used as the separator, e.g. 1k:4k. If the option allows two sets of ranges, they can be specified with a , or / delimiter: 1k-4k/8k-32k. Also see the int parameter type.
float_list — A list of floating point numbers, separated by a : character.

🌸 JOB PARAMETERS 🌸

With the above in mind, here follows the complete list of fio job parameters. Buckle up, bestie — this is the main character section 🫶

✨ Units ✨

kb_base=int — Select the interpretation of unit prefixes in input parameters.
- 1000 — Inputs comply with IEC 80000-13 and the International System of Units (SI). Use:
  - power-of-2 values with IEC prefixes (e.g., KiB)
  - power-of-10 values with SI prefixes (e.g., kB)
- 1024 — Compatibility mode (default). To avoid breaking old scripts:
  - power-of-2 values with SI prefixes
  - power-of-10 values with IEC prefixes

See bs for more details on input parameters.

Outputs always use correct prefixes. Most outputs include both side-by-side, like:

bw=2383.3kB/s (2327.4KiB/s)

If only one value is reported, then kb_base selects the one to use:

1000 — SI prefixes
1024 — IEC prefixes

unit_base=int — Base unit for reporting. Allowed values are:
- 0 — Use auto-detection (default).
- 8 — Byte based.
- 1 — Bit based.

🎀 Job description 🎀

name=str — ASCII name of the job. This may be used to override the name printed by fio for this job. Otherwise the job name is used. On the command line, this parameter has the special purpose of also signaling the start of a new job.
description=str — Text description of the job. Doesn’t do anything except dump this text description when this job is run. It’s not parsed.
loops=int — Run the specified number of iterations of this job. Used to repeat the same workload a given number of times. Defaults to 1.
numjobs=int — Create the specified number of clones of this job. Each clone of job is spawned as an independent thread or process. May be used to setup a larger number of threads/processes doing the same thing. Each thread is reported separately; to see statistics for all clones as a whole, use group_reporting in conjunction with new_group. See --max-jobs. Default: 1.

runtime=time — Limit runtime. The test will run until it completes the configured I/O workload or until it has run for this specified amount of time, whichever occurs first. It can be quite hard to determine for how long a specified job will run, so this parameter is handy to cap the total runtime to a given time. When the unit is omitted, the value is interpreted in seconds.
time_based — If set, fio will run for the duration of the runtime specified even if the file(s) are completely read or written. It will simply loop over the same workload as many times as the runtime allows.
startdelay=irange(int) — Delay the start of job for the specified amount of time. Can be a single value or a range. When given as a range, each thread will choose a value randomly from within the range. Value is in seconds if a unit is omitted.
ramp_time=time — If set, fio will run the specified workload for this amount of time before logging any performance numbers. Useful for letting performance settle before logging results, thus minimizing the runtime required for stable results. Note that the ramp_time is considered lead in time for a job, thus it will increase the total runtime if a special timeout or runtime is specified. When the unit is omitted, the value is given in seconds.
ramp_size=size — If set, fio will wait until the job does given amount of IO before logging any performance numbers. When group_reporting is enabled, the logging starts when all jobs in the group together perform given amount of IO. Similarly to ramp_time this is useful for letting performance to settle before logging results and will increase the total runtime if a special timeout or runtime is specified. When the unit is omitted, the value is given in bytes.
clocksource=str — Use the given clocksource as the base of timing. The supported options are:
- gettimeofday — gettimeofday(2)
- clock_gettime — clock_gettime(2)
- cpu — Internal CPU clock source

cpu is the preferred clocksource if it is reliable, as it is very fast (and fio is heavy on time calls). Fio will automatically use this clocksource if it’s supported and considered reliable on the system it is running on, unless another clocksource is specifically set. For x86/x86-64 CPUs, this means supporting TSC Invariant.

gtod_reduce=bool — Enable all of the gettimeofday(2) reducing options (disable_clat, disable_slat, disable_bw_measurement) plus reduce precision of the timeout somewhat to really shrink the gettimeofday(2) call count. With this option enabled, we only do about 0.4% of the gettimeofday(2) calls we would have done if all time keeping was enabled.
gtod_cpu=int — Sometimes it’s cheaper to dedicate a single thread of execution to just getting the current time. Fio (and databases, for instance) are very intensive on gettimeofday(2) calls. With this option, you can set one CPU aside for doing nothing but logging current time to a shared memory location. Then the other threads/processes that run I/O workloads need only copy that segment, instead of entering the kernel with a gettimeofday(2) call. The CPU set aside for doing these time calls will be excluded from other uses. Fio will manually clear it from the CPU mask of other jobs.
job_start_clock_id=int — The clock_id passed to the call to clock_gettime used to record job_start in the json output format. Default is 0, or CLOCK_REALTIME.

🎯 Target file/device 🎯

directory=str — this is your "put the files over there, not here" knob 💅 It prefixes filenames with this directory, so your files land somewhere other than ./. You can hand it a whole bunch of directories by separating the names with a : character, and those directories get assigned equally distributed to the job clones created by numjobs — as long as those clones are using generated filenames, that is. If specific filename(s) are set, fio just uses the first listed directory, and that way it matches the filename semantic (which generates a file for each clone if not specified, but lets all the clones use the same file if set). 🫶

Peep the filename option for how to escape : characters inside the directory path itself, iykyk.

Note: if you want to control the directory fio uses for its internal state files, that's --aux-path, bestie. Different thing entirely ✨

filename=str — okay so normally fio just makes up a filename based on the job name, thread number and file number (see filename_format). But if you want to share files between threads in a job, or across several jobs with fixed file paths, you specify a filename for each of them and that overrides the default. If the ioengine is file based, you can specify a number of files by separating the names with a : colon. So if you wanted a job to open /dev/sda and /dev/sdb as its two working files, you'd do filename=/dev/sda:/dev/sdb — iconic, honestly. This also means that whenever this option is specified, nrfiles is ignored. The size of regular files specified by this option will be size divided by number of files, unless an explicit size is specified by filesize.

Each colon in the wanted path must be escaped with a \ character. For instance, if the path is /dev/dsk/foo@3,0:c then you would use filename=/dev/dsk/foo@3,0\:c, and if the path is F:\filename then you would use filename=F\:\filename. Escape or perish 😭

On Windows, disk devices are accessed as \\.\PhysicalDrive0 for the first device, \\.\PhysicalDrive1 for the second etc.

Note: Windows and FreeBSD prevent write access to areas of the disk containing in-use data (e.g. filesystems). So like… be normal about it 💀

For HTTP and S3 access, specify a valid URL path or S3 key, respectively. A filename for path-style S3 includes a bucket name (/bucket/k/e.y) while a virtual-hosted-style S3 filename (/k/e.y) does not, because its bucket name is specified in http_host.

The filename - is a reserved name, meaning stdin or stdout. Which of the two depends on the read/write direction set. Sneaky but slay 🌈

filename_format=str — if you're sharing multiple files between jobs, you usually need fio to generate the exact names you want. By default, fio names a file based on the default file format specification of jobname.jobnumber.filenumber. With this option, that whole situation is customizable. Fio will recognize and replace the following keywords in this string:
- $jobname — the name of the worker thread or process.
- $clientuid — IP of the fio process when using client/server mode.
- $jobnum — the incremental number of the worker thread or process.
- $filenum — the incremental number of the file for that worker thread or process.

To have dependent jobs share a set of files, this option can be set to have fio generate filenames that are shared between the two. For instance, if testfiles.$filenum is specified, file number 4 for any job will be named testfiles.4. The default of $jobname.$jobnum.$filenum will be used if no other format specifier is given. 🎀

If you specify a path then the directories will be created up to the main directory for the file. So for example, if you specify a/b/c/$jobnum then the directories a/b/c will be created before the file setup part of the job. If you specify directory then the path will be relative to that directory, otherwise it is treated as the absolute path. Path brainrot, but the good kind ✨

unique_filename=bool — to avoid collisions between networked clients, fio defaults to prefixing any generated filenames (with a directory specified) with the source of the client connecting. To disable this behavior, set this option to 0. Lowkey a lifesaver in client/server mode, ngl 🦋
filetype=str — assume that all files defined in a job are of this type. By default fio will do stat(2) for each file to know its file type. For huge filesets that might be a bottleneck, so this option can be used to skip the huge number of syscalls. The file types are:
- none — unset. The default.
- file — regular file.
- block — block device file.
- char — char device file.
opendir=str — recursively open any files below directory str. This accepts only a single directory and, unlike related options, colons appearing in the path must not be escaped. Not me giving you two opposite colon rules in one page 🙈 that's just how it is, bestie.
lockfile=str — fio defaults to not locking any files before it does I/O to them. If a file or file descriptor is shared, fio can serialize I/O to that file to make the end result consistent. This is usual for emulating real workloads that share files. The lock modes are:
- none — no locking. The default.
- exclusive — only one thread or process may do I/O at a time, excluding all others.
- readwrite — read-write locking on the file. Many readers may access the file at the same time, but writes get exclusive access.
nrfiles=int — number of files to use for this job. Defaults to 1. The size of files will be size divided by this, unless explicit size is specified by filesize. Files are created for each thread separately, and each file will have a file number within its name by default, as explained in the filename section. Math girl summer 🔥
openfiles=int — number of files to keep open at the same time. Defaults to the same as nrfiles, and can be set smaller to limit the number simultaneous opens. Efficiency era 💅
file_service_type=str — defines how fio decides which file from a job to service next. The following types are defined:
- random — choose a file at random.
- roundrobin — round robin over opened files. This is the default.
- sequential — finish one file before moving on to the next. Multiple files can still be open depending on openfiles.
- zipf — use a Zipf distribution to decide what file to access.
- pareto — use a Pareto distribution to decide what file to access.
- normal — use a Gaussian (normal) distribution to decide what file to access.
- gauss — alias for normal.

For random, roundrobin and sequential, a postfix can be appended to tell fio how many I/Os to issue before switching to a new file. For example, specifying file_service_type=random:8 would cause fio to issue 8 I/Os before selecting a new file at random. For the non-uniform distributions, a floating point postfix can be given to influence how the distribution is skewed. See random_distribution for a description of how that would work. Distribution lore 🧚

ioscheduler=str — attempt to switch the device hosting the file to the specified I/O scheduler before running. If the file is a pipe, a character device file, or if the device hosting the file could not be determined, this option is ignored. It's giving "nice try though" 🌸
create_serialize=bool — if true, serialize the file creation for the jobs. This may be handy to avoid interleaving of data files, which may greatly depend on the filesystem used and even the number of processors in the system. Default: true. Chaos-prevention queen 👑
create_fsync=bool — fsync(2) the data file after creation. This is the default. Flush it, bestie 💖
create_on_open=bool — if true, don't pre-create files but allow the job's open() to create a file when it's time to do I/O. Default: false — i.e. pre-create all necessary files when the job starts.
create_only=bool — if true, fio will only run the setup phase of the job. If files need to be laid out or updated on disk, only that will be done — the actual job contents are not executed. Default: false. All vibes, no workload 😌
allow_file_create=bool — if true, fio is permitted to create files as part of its workload. If this option is false, then fio will error out if the files it needs to use don't already exist. Default: true. Strict mode activated 🫶
allow_mounted_write=bool — if this isn't set, fio will abort jobs that are destructive (e.g. that write) to what appears to be a mounted device or partition. This should help catch creating inadvertently destructive tests, not realizing that the test will destroy data on the mounted file system. Note that some platforms don't allow writing against a mounted device regardless of this option. Default: false.

ok bestie this one is serious — it's literally there so you don't nuke your data by accident. Genuinely do not go flipping this on without knowing what you're doing 😭

pre_read=bool — if this is given, files will be pre-read into memory before starting the given I/O operation. This will also clear the invalidate flag, since it is pointless to pre-read and then drop the cache. This will only work for I/O engines that are seekable, since they allow you to read the same data multiple times. Thus, it will not work on non-seekable I/O engines (e.g. network, splice). Default: false. Read-ahead slay, but only for seekable engines 💀
unlink=bool — unlink (delete) the job files when done. Not the default, as repeated runs of that job would then waste time recreating the file set again and again. Default: false. Cleanup crew vs. efficiency, pick your fighter 🦋
unlink_each_loop=bool — unlink (delete) job files after each iteration or loop. Default: false. Absolutely feral behavior, but it's there if you want it 🍓
zonemode=str — accepted values are:
- none — the zonerange, zonesize, zonecapacity and zoneskip parameters are ignored.
strided — bestie, this one's a whole vibe: I/O happens inside one single zone until zonesize bytes have been transferred. The second that many bytes are done, processing of the next zone kicks off. And zonecapacity? Yeah, it just gets ignored here. 💅
zbd — Zoned block device mode, the main character 🧚. I/O happens sequentially in each zone, even if you selected random I/O. Random I/O happens across all zones instead of being restricted to a single zone. The zoneskip parameter is ignored. zonerange and zonesize must be identical. Here's the tea though 😭: if the size of the unwritten space in a zone isn't a multiple of the specified block size at workload start, write workloads create unwritten remainder areas at the ends of zones and keep the zones in open conditions. To free up the open zone resources, fio issues zone finish operations to the zones with the remainders. Trim is handled using a zone reset operation. Trim only considers non-empty sequential write required and sequential write preferred zones.
zonerange=int — For zonemode=strided, this is the size of a single zone (peep zonesize and zoneskip too 🫶). For zonemode=zbd, this parameter is ignored, sorry babe.
zonesize=int — For zonemode=strided, this is the number of bytes to transfer before skipping zoneskip bytes. If this parameter is smaller than zonerange then only a fraction of each zone with zonerange bytes will be accessed. If it's larger than zonerange then each zone will be accessed multiple times before skipping to the next zone 👀. For zonemode=zbd, this is the size of a single zone, and zonerange is ignored in this mode. For a job accessing a zoned block device, the specified zonesize must be 0 or equal to the device zone size. For a regular block device or file, the specified zonesize must be at least 512B.
zonecapacity=int — For zonemode=zbd, this defines the capacity of a single zone, which is the accessible area starting from the zone start address ✨. This parameter only applies when using zonemode=zbd in combination with regular block devices. If not specified it defaults to the zone size. If the target device is a zoned block device, the zone capacity is obtained from the device information and this option is ignored 💀.
zoneskip=int[z] — For zonemode=strided, the number of bytes to skip after zonesize bytes of data have been transferred. For zonemode=zbd, the zonesize aligned number of bytes to skip once a zone is fully written (write workloads) or all written data in the zone have been read (read workloads). This parameter is valid only for sequential workloads and ignored for random workloads. For read workloads, see also read_beyond_wp 👑
read_beyond_wp=bool — This parameter applies to zonemode=zbd only.

Ok so zoned block devices are block devices made of multiple zones, right? Each zone has a type, e.g. conventional or sequential. A conventional zone can be written at any offset that's a multiple of the block size. Sequential zones must be written sequentially — no skipping the line, periodt. The position at which a write must occur is called the write pointer 🎯. A zoned block device can be either host managed or host aware. For host managed devices the host must ensure that writes happen sequentially. Fio recognizes host managed devices and serializes writes to sequential zones for these devices.

So if a read happens in a sequential zone beyond the write pointer, the zoned block device will complete the read without reading any data from the storage medium 😭😭. And since such reads lead to unrealistically high bandwidth and IOPS numbers, fio only reads beyond the write pointer if you explicitly tell it to. Default: false.

max_open_zones=int — When a zone of a zoned block device is partially written (i.e. not all sectors of the zone have been written), the zone is in one of three conditions: 'implicit open', 'explicit open' or 'closed'. Zoned block devices may have a limit called 'max_open_zones' (same name as the parameter, cute right 💖) on the total number of zones that can simultaneously be in the 'implicit open' or 'explicit open' conditions. Zoned block devices may have another limit called 'max_active_zones', on the total number of zones that can simultaneously be in the three conditions. The max_open_zones parameter limits the number of zones to which write commands are issued by all fio jobs — that is, it limits the number of zones that will be in the conditions. When the device has the max_open_zones limit and does not have the max_active_zones limit, the max_open_zones parameter limits the number of zones in the two open conditions up to the limit. In this case, fio includes zones in the two open conditions to the write target zones at fio start. When the device has both the max_open_zones and the max_active_zones limits, the max_open_zones parameter limits the number of zones in the three conditions up to the limit. In this case, fio includes zones in the three conditions to the write target zones at fio start. 🦋

This parameter is relevant only if zonemode=zbd is used. The default value is always equal to the max_open_zones limit of the target zoned block device, and a value higher than this limit cannot be specified by users unless the option ignore_zone_limits is specified. When ignore_zone_limits is specified, or the target device does not have the max_open_zones limit, max_open_zones can specify 0 to disable any limit on the number of zones that can be simultaneously written to by all jobs 🙈.

job_max_open_zones=int — Same energy as max_open_zones, but it limits the number of open zones per fio job — i.e. the number of zones a single job can simultaneously write to. A value of zero means no limit. Default: zero.
ignore_zone_limits=bool — If this option is used, fio will ignore the maximum number of open zones limit of the zoned block device in use, thus allowing the option max_open_zones value to be larger than the device reported limit 🔥. Default: false.
zone_reset_threshold=float — A number between zero and one that indicates the ratio of written bytes in the zones with write pointers in the IO range to the size of the IO range. When the current ratio is above this ratio, zones are reset periodically as zone_reset_frequency specifies. If there are multiple jobs when using this option, the IO range for all write jobs has to be the same. No exceptions 💅
zone_reset_frequency=float — A number between zero and one that indicates how often a zone reset should be issued if the zone reset threshold has been exceeded. A zone reset is submitted after each (1 / zone_reset_frequency) write requests. This and the previous parameter can be used to simulate garbage collection activity ✨ iconic honestly.
recover_zbd_write_error=bool — If this option is specified together with the option continue_on_error, check the write pointer positions after the failed writes to sequential write required zones. Then move the write pointers so that the next writes don't fail due to partial writes and unexpected write pointer positions 🫶. If continue_on_error is not specified, it errors out 💀. When the writes are asynchronous, the write pointer move fills blocks with zero and then breaks verify data. If an asynchronous IO engine and verify workload are specified, it errors out. Default: false.
write_zone_remainders=bool — If the size of the unwritten space in a zone is not a multiple of the specified block size at workload start, write workloads create unwritten remainder areas at the ends of zones. By default, fio issues zone finish operations on such zones, transitioning them to the full condition and freeing up open zone resources. However, zone finish operations introduce waits for in-flight writes, reducing overall write throughput 😭. If this option is specified, fio writes data to the remainder areas instead of performing zone finish operations. This improves write throughput by avoiding waits for in-flight writes — particularly in asynchronous write workloads ✨. The drawback of this option is that it requires fio to perform writes smaller than the minimum block size. Consequently, the option norandommap must be set. If norandommap is not set, it is automatically set. Default: false.

🌈 I/O type 🌈

direct=bool — If value is true, use non-buffered I/O. This is usually O_DIRECT. Note that OpenBSD and ZFS on Solaris don't support direct I/O 🙈. On Windows the synchronous ioengines don't support direct I/O. Default: false.
buffered=bool — If value is true, use buffered I/O. This is the exact opposite of the direct option, bestie 💖. Defaults to true.
readwrite=str, rw=str — The type of I/O pattern. Accepted values are:
- read — Sequential reads.
- write — Sequential writes.
- trim — Sequential trims (Linux block devices and SCSI character devices only ⚠️).
- randread — Random reads.
- randwrite — Random writes.
- randtrim — Random trims (Linux block devices and SCSI character devices only ⚠️).
- rw, readwrite — Sequential mixed reads and writes 👑.
- randrw — Random mixed reads and writes 🔥.

`trimwrite`

Sequential trim+write sequences, bestie. Blocks get trimmed first, then those exact same blocks get written to. So if you set io_size=64K, Fio will trim a total of 64K bytes and write 64K bytes on the same trimmed blocks. This behaviour stays consistent with number_ios or any other Fio options that limit the total bytes or the number of I/O's. 💅

`randtrimwrite`

Like trimwrite, but it uses random offsets instead of sequential writes. Same energy, just chaotic. 🎲

Fio defaults to read if you don't specify the option, periodt. For the mixed I/O types, the default is to split them 50/50. For certain types of I/O the result may still be skewed a bit, since the speed may be different — no cap, that's just how it is. 🙈

You can also specify the number of I/Os to do before getting a new offset by appending :<nr> to the end of the string you give. For a random read, it would look like rw=randread:8 for passing in an offset modifier with a value of 8. If the suffix is used with a sequential I/O pattern, then the <nr> value you specified gets added to the generated offset for each I/O, turning sequential I/O into sequential I/O with holes. For instance, using rw=write:4k will skip 4k for every write. Also peep the rw_sequencer option. If this is used with verify then the verify_header_seed option will be disabled, unless it's explicitly enabled. ⭐

`rw_sequencer=str`

If an offset modifier is given by appending a number to the rw=str line, then this option controls how that number modifies the I/O offset being generated. Accepted values are:

sequential — generate a sequential offset.
identical — generate the same offset.

sequential is only useful for random I/O, where fio would normally generate a new random offset for every I/O. If you append e.g. 8 to randread, i.e. rw=randread:8, you would get a new random offset for every 8 I/Os. The result would be a sequence of 8 sequential offsets with a random starting point. However this behavior may change if a sequential I/O reaches end of the file. As sequential I/O is already sequential, setting sequential for that would not result in any difference. identical behaves in a similar fashion, except it sends the same offset 8 number of times before generating a new offset. 🦋

Example #1:

rw=randread:8
rw_sequencer=sequential
bs=4k

The generated sequence of offsets will look like this: 4k, 8k, 12k, 16k, 20k, 24k, 28k, 32k, 92k, 96k, 100k, 104k, 108k, 112k, 116k, 120k, 48k, 52k ...

Example #2:

rw=randread:8
rw_sequencer=identical
bs=4k

The generated sequence of offsets will look like this: 4k, 4k, 4k, 4k, 4k, 4k, 4k, 4k, 92k, 92k, 92k, 92k, 92k, 92k, 92k, 92k, 48k, 48k, 48k ...

`unified_rw_reporting=str`

Fio normally reports statistics on a per data direction basis — meaning reads, writes, and trims are accounted and reported separately. This option determines whether fio reports the results normally, summed together, or as both options. Accepted values are:

none — normal statistics reporting.
mixed — statistics are summed per data direction and reported together.
both — statistics are reported normally, followed by the mixed statistics.
0 — backward-compatible alias for none.
1 — backward-compatible alias for mixed.
2 — alias for both.

`randrepeat=bool`

Seed all random number generators in a predictable way so the pattern is repeatable across runs. Default: true. 🔁

`allrandrepeat=bool`

Alias for randrepeat. Default: true.

`randseed=int`

Seed the random number generators based on this seed value, so you can control what sequence of output is being generated. If not set, the random sequence depends on the randrepeat setting. 🎀

`fallocate=str`

Whether pre-allocation is performed when laying down files. Accepted values are:

none — do not pre-allocate space.
native — use a platform's native pre-allocation call but fall back to none behavior if it fails/is not implemented.
posix — pre-allocate via posix_fallocate(3).
keep — pre-allocate via fallocate(2) with FALLOC_FL_KEEP_SIZE set.
truncate — extend file to final size using ftruncate(2) instead of allocating.
0 — backward-compatible alias for none.
1 — backward-compatible alias for posix.

May not be available on all supported platforms. keep is only available on Linux. If using ZFS on Solaris this cannot be set to posix because ZFS doesn't support pre-allocation. Default: native if any pre-allocation methods except truncate are available, none if not. 👑

Ngl, using truncate on Windows will interact surprisingly with non-sequential write patterns. When writing to a file that has been extended by setting the end-of-file information, Windows will backfill the unwritten portion of the file up to that offset with zeroes before issuing the new write. This means that a single small write to the end of an extended file will stall until the entire file has been filled with zeroes. 😭

`fadvise_hint=str`

Use posix_fadvise(2) or posix_madvise(2) to advise the kernel what I/O patterns are likely to be issued. Accepted values are:

0 — backwards compatible hint for "no hint".
1 — backwards compatible hint for "advise with fio workload type". This uses FADV_RANDOM for a random workload, and FADV_SEQUENTIAL for a sequential workload.
sequential — advise using FADV_SEQUENTIAL.
random — advise using FADV_RANDOM.
noreuse — advise using FADV_NOREUSE. This may be a no-op on older Linux kernels. Since Linux 6.3, it provides a hint to the LRU algorithm. See the posix_fadvise(2) man page. 💜

`write_hint=str`

Use fcntl(2) to advise the kernel what life time to expect from a write. Only supported on Linux, as of version 4.13. Accepted values are:

none — no particular life time associated with this file.
short — data written to this file has a short life time.
medium — data written to this file has a medium life time.
long — data written to this file has a long life time.
extreme — data written to this file has a very long life time.

The values are all relative to each other, and no absolute meaning should be associated with them. ✨

`offset=int[%|z]`

Start I/O at the provided offset in the file, given as either a fixed size in bytes, zones or a percentage. If a percentage is given, the generated offset will be aligned to the minimum blocksize or to the value of offset_align if provided. Data before the given offset will not be touched. This effectively caps the file size at real_size - offset. Can be combined with size to constrain the start and end range of the I/O workload. A percentage can be specified by a number between 1 and 100 followed by %, for example, offset=20% to specify 20%. In ZBD mode, value can be set as number of zones using z. 🌸

`offset_align=int`

If set to a non-zero value, the byte offset generated by a percentage offset is aligned upwards to this value. Defaults to 0, meaning that a percentage offset is aligned to the minimum block size.

`offset_increment=int[%|z]`

If this is provided, then the real offset becomes offset + offset_increment * thread_number, where the thread number is a counter that starts at 0 and is incremented for each sub-job (i.e. when numjobs option is specified). This option is useful if there are several jobs which are intended to operate on a file in parallel disjoint segments, with even spacing between the starting points. Percentages can be used for this option. If a percentage is given, the generated offset will be aligned to the minimum blocksize or to the value of offset_align if provided. In ZBD mode, value can be set as number of zones using z. 🧚

`number_ios=int`

Fio will normally perform I/Os until it has exhausted the size of the region set by size, or if it exhausts the allocated time (or hits an error condition). With this setting, the range/size can be set independently of the number of I/Os to perform. When fio reaches this number, it will exit normally and report status. Note that this does not extend the amount of I/O that will be done — it will only stop fio if this condition is met before other end-of-job criteria. 💀

`fsync=int`

If writing to a file, issue an fsync(2) (or its equivalent) of the dirty data for every number of blocks given. For example, if you give 32 as a parameter, fio will sync the file after every 32 writes issued. If fio is using non-buffered I/O, we may not sync the file. The exception is the sg I/O engine, which synchronizes the disk cache anyway. Defaults to 0, which means fio does not periodically issue and wait for a sync to complete. Also see end_fsync, end_syncfs and fsync_on_close. 🔥

`fdatasync=int`

Like fsync but uses fdatasync(2) to only sync data and not metadata blocks. In Windows, DragonFlyBSD or OSX there is no fdatasync(2) so this falls back to using fsync(2). Defaults to 0, which means fio does not periodically issue and wait for a data-only sync to complete. 🫶

`write_barrier=int`

Make every N-th write a barrier write. 🚧

sync_file_range=str:int — makes fio call sync_file_range(2) every int number of write ops. fio keeps track of the range of writes that have piled up since the last sync_file_range(2) call, and str can currently be one or more of:
- wait_before — SYNC_FILE_RANGE_WAIT_BEFORE
- write — SYNC_FILE_RANGE_WRITE
- wait_after — SYNC_FILE_RANGE_WRITE_AFTER

So if you do sync_file_range=wait_before,write:8, fio would use SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE for every 8 writes. Also see the sync_file_range(2) man page. Heads up bestie, this option is Linux specific.

overwrite=bool — if true, writes to a file will always overwrite existing data. If the file doesn't already exist, it gets created before the write phase begins. And if the file exists and is already large enough for the specified write phase, literally nothing will be done. Default: false.
end_fsync=bool — if true, fsync(2) the file contents once a write stage has completed. Default: false.
end_syncfs=bool — same energy as end_fsync, except instead of executing fsync(2) for each file of a write stage, it executes syncfs(2) to synchronize all the written files with a single system call when a write stage has completed. Default: false.
fsync_on_close=bool — if true, fio will fsync(2) a dirty file on close. The difference from end_fsync is that this happens on every file close, not just at the end of the job. Default: false.
rwmixread=int — the percentage of a mixed workload that should be reads. Default: 50.
rwmixwrite=int — the percentage of a mixed workload that should be writes. If both rwmixread and rwmixwrite are given and the values don't add up to 100%, the latter of the two gets used to override the first. Fair warning: this may interfere with a given rate setting if fio is asked to limit reads or writes to a certain rate — and if that's the case, the distribution may be skewed. Default: 50.
random_distribution=str:float[:float][,str:float][,str:float] — by default, fio uses a completely uniform random distribution when asked to perform random I/O. Sometimes it's useful to skew the distribution in specific ways, so some parts of the data are more hot than others 🔥. fio includes these distribution models:
- random — uniform random distribution
- zipf — Zipf distribution
- pareto — Pareto distribution
- normal — Normal (Gaussian) distribution
- zoned — Zoned random distribution
- zoned_abs — Zoned absolute random distribution

When you're using a zipf or pareto distribution, an input value is also needed to define the access pattern. For zipf, that's the Zipf theta. For pareto, it's the Pareto power. fio includes a test program, fio-genzipf, that you can use to visualize what the given input values will yield in terms of hit rates. If you wanted to use zipf with a theta of 1.2, you'd use random_distribution=zipf:1.2 as the option. And if a non-uniform model is used, fio will disable use of the random map. For the normal distribution, a normal (Gaussian) deviation is supplied as a value between 0 and 100.

The second, optional float is allowed for pareto, zipf and normal distributions. It lets you set the base of the distribution in a non-default place, which gives you more control over the most probable outcome. This value is in the range [0-1], which maps linearly to the range of possible random values. Defaults are: random for pareto and zipf, and 0.5 for normal. So if you wanted to use zipf with a theta of 1.2 centered on 1/4 of the allowed value range, you'd use random_distribution=zipf:1.2:0.25. Slay.

For a zoned distribution, fio supports specifying percentages of I/O access that should fall within what range of the file or device. For example, given a criteria of:

60% of accesses should be to the first 10%
30% of accesses should be to the next 20%
8% of accesses should be to the next 30%
2% of accesses should be to the next 40%

we can define that through zoning of the random accesses. For the above example, the user would do:

random_distribution=zoned:60/10:30/20:8/30:2/40

A zoned_abs distribution works exactly like the zoned, except that it takes absolute sizes. For example, let's say you wanted to define access according to the following criteria:

60% of accesses should be to the first 20G
30% of accesses should be to the next 100G
10% of accesses should be to the next 500G

we can define an absolute zoning distribution with:

random_distribution=zoned:60/10:30/20:8/30:2/40

For both zoned and zoned_abs, fio supports defining up to 256 separate zones.

Similarly to how bssplit works for setting ranges and percentages of block sizes. Like bssplit, it's possible to specify separate zones for reads, writes, and trims. If just one set is given, it'll apply to all of them 💅.

percentage_random=int[,int][,int] — for a random workload, sets how big a percentage should be random. This defaults to 100%, in which case the workload is fully random. It can be set from anywhere from 0 to 100. Setting it to 0 would make the workload fully sequential. Any setting in between gives you a random mix of sequential and random I/O, at the given percentages. Comma-separated values may be specified for reads, writes, and trims as described in blocksize.
norandommap — normally fio will cover every block of the file when doing random I/O. If this option is given, fio just gets a new random offset without looking at past I/O history. That means some blocks may not be read or written, and some blocks may be read/written more than once 😭. If this option is used with verify then verify_header_seed will be disabled. If this option is used with verify and multiple blocksizes (via bsrange), only intact blocks are verified, i.e., partially-overwritten blocks are ignored. With an async I/O engine and an I/O depth > 1, header write sequence number verification will be disabled. See verify_write_sequence.
softrandommap=bool — see norandommap. If fio runs with the random block map enabled and it fails to allocate the map, if this option is set it will continue without a random block map. Since coverage will not be as complete as with random maps, this option is disabled by default.
random_generator=str — fio supports these engines for generating I/O offsets for random I/O:
- tausworthe — strong 2^88 cycle random number generator.
- lfsr — linear feedback shift register generator.
- tausworthe64 — strong 64-bit 2^258 cycle random number generator.

tausworthe is a strong random number generator, but it requires tracking on the side if we want to ensure blocks are only read or written once. lfsr guarantees that we never generate the same offset twice, and it's also less computationally expensive. It's not a true random generator, however, though for I/O purposes it's typically good enough. lfsr only works with single block sizes, not with workloads that use multiple block sizes — if used with such a workload, fio may read or write some blocks multiple times 🙈. The default value is tausworthe, unless the required space exceeds 2^32 blocks. If it does, then tausworthe64 is selected automatically.

sprandom=bool — SPRandom is a method designed to rapidly precondition SSDs for steady-state random write workloads. It divides the device into equally sized regions and writes the device's entire physical capacity once, selecting offsets so that the regions have a distribution of invalid blocks matching the distribution that occurs at steady state. Default: false.

It uses random_generator=lfsr, which fio will set by default. Selecting any other random generator will result in an error ⚠️.

spr_num_regions=int — see sprandom. Specifies the number of regions used for SPRandom. Default=100.

For large devices it's better to use more regions, to increase precision and reduce memory allocation. The allocation is proportional to the region size.

spr_cs=int — see sprandom. Defines a cache size in bytes, as specified by the SSD manufacturer.

When this is non-zero, delay invalidating writes by one region in order to make sure that all original writes from a region are flushed from cache before the later invalidating writes are sent to the device. This deferral prevents the original write and the later invalidating write from being present in the device's cache at the same time — which would allow the device to ignore the original write and prevent sprandom from achieving its target validity fractions 🎯. The actual cache size is used to ensure the number of regions isn't set so large that the size of a region is smaller than the device cache. The default is 0.

spr_op=float — see sprandom. Over-provisioning ratio in the range (0, 1), as specified by the SSD manufacturer. The default is 0.15.

🧱 Block size 🧱

blocksize=int[,int][,int], bs=int[,int][,int] — the block size in bytes used for I/O units. Default: 4096. A single value applies to reads, writes, and trims. Comma-separated values may be specified for reads, writes, and trims. A value not terminated in a comma applies to subsequent types. Examples, bestie:

bs=256k        means 256k for reads, writes and trims.
bs=8k,32k      means 8k for reads, 32k for writes and trims.
bs=8k,32k,     means 8k for reads, 32k for writes, and default for trims.
bs=,8k         means default for reads, 8k for writes and trims.
bs=,8k,        means default for reads, 8k for writes, and default for trims.

See how the trailing comma is doing so much work? Iconic ✨.

blocksize_range=irange[,irange][,irange], bsrange=irange[,irange][,irange] — bestie, this is how you say "the block sizes I want are somewhere in this range" (in bytes, for I/O units). The I/O unit that actually gets issued will always be a multiple of the minimum size — unless you've set blocksize_unaligned, which we'll get to. You can also give comma-separated ranges for reads, writes and trims, exactly like blocksize does. Example time ✨:

bsrange=1k-4k,2k-8k or bsrange=1k:4k,2k:8k

bssplit=str[,str][,str] — okay so sometimes a plain even split between block sizes just isn't the vibe and you need finer grained control 👑. This option lets you weight the block sizes so you can declare exactly how much of each size gets issued. The format is literally:

bssplit=blocksize/percentage:blocksize/percentage

for as many block sizes as you need. So if your dream workload is 50% 64k blocks, 10% 4k blocks and 40% 32k blocks, you'd write:

bssplit=4k/10:64k/50:32k/40

Ordering? Doesn't matter, it's not that deep 💅. If you leave a percentage blank, fio fills in the rest evenly. So a bssplit like:

bssplit=4k/50:1k/:32k/

gives you 50% 4k ios, and 25% each of 1k and 32k ios. The percentages always add up to 100 — if you hand it a range that adds up to more, it will error out, no cap 😭.

Comma-separated values may be specified for reads, writes and trims as described in blocksize.

Wanna do 50% 2k reads + 50% 4k reads, while also having 90% 4k writes + 10% 8k writes? You'd specify:

bssplit=2k/50:4k/50,4k/90:8k/10

Fio supports defining up to 64 different weights for each data direction. Sixty-four!! 🫶

blocksize_unaligned, bs_unaligned — if set, fio will issue I/O units with any size within blocksize_range, not just multiples of the minimum size. Heads up though: this typically won't work with direct I/O, because direct I/O normally requires sector alignment. Not me warning you in advance 🙈.
bs_is_seq_rand=bool — set this and fio flips the script: your normal read,write blocksize settings get used as sequential,random blocksize settings instead. So any random read or write uses the WRITE blocksize settings, and any sequential read or write uses the READ blocksize settings. The way it just swaps them 🔁.
blockalign=int[,int][,int], ba=int[,int][,int] — the boundary that random I/O units get aligned to. Default: blocksize. Minimum alignment is typically 512b for direct I/O, though honestly it usually depends on your hardware block size. This option is mutually exclusive with using a random map for files, so setting it will turn that option off. Comma-separated values may be specified for reads, writes and trims as described in blocksize.

🧺 Buffers and memory 🧺

zero_buffers — initialize the buffers with all zeros. Default is that fio fills buffers with random data. Zeroes are also a vibe, just a quieter one 🌸.
refill_buffers — give this option and fio refills the I/O buffers on every single submit. Default behaviour is to only fill them at init time and then reuse that data forever. Only makes sense if zero_buffers isn't specified, naturally. And hey — if data verification is enabled, refill_buffers is automatically enabled too. It said "I'll handle it" 💅.
scramble_buffers=bool — if refill_buffers is too costly and your target is doing data deduplication, set this and fio will slightly modify the I/O buffer contents to defeat normal de-dupe attempts. Lowkey important: this is not enough to defeat more clever block compression attempts, but it will stop naive dedupe of blocks. Default: true.
buffer_compress_percentage=int — set this and fio tries to give you I/O buffer content (on WRITEs) that compresses to that specified level, by providing a mix of random data followed by fixed pattern data. The fixed pattern is either zeros or whatever buffer_pattern says — and if buffer_pattern is used, it might skew the compression ratio slightly, fyi. Setting buffer_compress_percentage to a value other than 100 also enables refill_buffers, so adjacent blocks are less likely to be so similar that they over-compress when seen together. Peep buffer_compress_chunk for how to set a finer or coarser granularity of the random/fixed data regions. Defaults to unset — i.e. buffer data won't adhere to any compression level at all.
buffer_compress_chunk=int — this one lets fio manage how big that random/fixed data region is when you're using buffer_compress_percentage. When buffer_compress_chunk is set to some non-zero value smaller than the block size, fio can repeat the random/fixed region throughout the I/O buffer at the specified interval (which is especially useful when bigger block sizes are used for a job 🔥). Set it to 0 and fio uses a chunk size matching the block size, giving you a single random/fixed region inside the I/O buffer. Defaults to 512. When the unit is omitted, the value is interpreted in bytes.
buffer_pattern=str — if set, fio fills the I/O buffers with this pattern, or with the contents of a file 🎀. If not set, buffer contents are defined by all the other buffer-content options. The setting can be any pattern of bytes, and can be prefixed with 0x for hex values. It may also be a string — in which case the string must be wrapped with "". Or it may be a filename — wrapped with '', in which case the file gets opened and read. Note that not all the file contents will be read if that would cause the buffers to overflow. So, for example:

buffer_pattern='filename'
or:
buffer_pattern="abcd"
or:
buffer_pattern=-12
or:
buffer_pattern=0xdeadface

Also, you can combine everything together in any order, icon behaviour honestly:

buffer_pattern=0xdeadface"abcd"-12'filename'

dedupe_percentage=int — if set, fio generates this percentage of identical buffers when writing. These buffers will be naturally dedupable ✨. The contents depend on whatever other buffer compression settings you've set — so individual buffers can be either fully compressible or not at all. This option only controls the distribution of unique buffers. Setting it also enables refill_buffers, so you don't end up with every buffer being identical.
dedupe_mode=str — when dedupe_percentage is given, this controls how fio generates the dedupe buffers.
- repeat — generate dedupe buffers by repeating previous writes.
- working_set — generate dedupe buffers from the working set.

repeat is the default for fio — dedupe buffers are generated by repeating a previous unique write.

working_set is a more realistic workload. With working_set, you should provide dedupe_working_set_percentage. Given that, fio uses the initial unique write buffers as its working set, and upon deciding to dedupe, it randomly chooses a buffer from the working set 🦋. Note that by using working_set the dedupe percentage will converge to the desired one over time, while repeat maintains the desired percentage throughout the whole job.

dedupe_working_set_percentage=int — if dedupe_mode is set to working_set, this controls the percentage of the size of the file or device used as the buffers fio picks the dedupe buffers from.

Note that size needs to be explicitly provided, and only 1 file per job is supported 🙈.

dedupe_global=bool — controls whether the deduplication buffers get shared amongst all jobs that have this option set. The buffers are spread evenly between participating jobs. Very "sharing is caring" of them 🫶.

Note that dedupe_mode must be set to working_set for this to work. Can be used in combination with compression.

invalidate=bool — invalidate the buffer/page cache parts of the files to be used prior to starting I/O, if the platform and file type support it. Defaults to true. This will be ignored if pre_read is also specified for the same job 👀.
sync=str — whether, and what type of, synchronous I/O to use for writes. The allowed values are:
- none — do not use synchronous IO, the default.
- 0 — same as none.
- sync — use synchronous file IO. For the majority of I/O engines this means using O_SYNC.
- 1 — same as sync.
- dsync — use synchronous data IO. For the majority of I/O engines this means using O_DSYNC.
iomem=str, mem=str — fio can use various types of memory as the I/O unit buffer. The allowed values are:
- malloc — use memory from malloc(3) as the buffers. Default memory type.
- shm — use shared memory as the buffers, allocated through shmget(2).
- shmhuge — same as shm, but use huge pages as backing.
- mmap — use mmap(2) to allocate buffers. May either be anonymous memory, or can be file backed if a filename is given after the option. The format is mem=mmap:/path/to/file.
- mmaphuge — use a memory mapped huge file as the buffer backing. Append the filename after mmaphuge, ala mem=mmaphuge:/hugetlbfs/file.
- mmapshared — same as mmap, but use a MMAP_SHARED mapping.
- cudamalloc — use GPU memory as the buffers for GPUDirect RDMA benchmark. The ioengine must be rdma 💀 (no but seriously, it has to be).

The area allocated is a function of the maximum allowed bs size for the job, multiplied by the I/O depth given. Real talk though — for shmhuge and mmaphuge to actually work, your system has to have free huge pages allocated already, no cap. You can normally check and set that by reading/writing /proc/sys/vm/nr_hugepages on a Linux system. Fio assumes a huge page is 2 or 4MiB in size depending on the platform. So to calculate the number of huge pages you need for a given job file: add up the I/O depth of all jobs (normally one unless iodepth is used) and multiply by the maximum bs set. Then divide that number by the huge page size. You can peep the size of the huge pages in /proc/meminfo. And heads up — if no huge pages are allocated (like, if nr_hugepages is a non-zero number... or rather, if it isn't), using mmaphuge or shmhuge will fail. Also see hugepage-size. 💖

mmaphuge also needs to have hugetlbfs mounted, and the file location should point there. So if it's mounted in /huge, you would use mem=mmaphuge:/huge/somefile. Easy ✨

iomem_align=int, mem_align=int — this is the memory alignment of the I/O memory buffers. Bestie, note the little catch: the given alignment is applied to the first I/O unit buffer, and if you're using iodepth, the alignment of the following buffers is given by the bs used. In other words — if you use a bs that's a multiple of the page size in the system, all buffers will be aligned to this value. If you use a bs that is not page aligned, the alignment of subsequent I/O memory buffers is the sum of the iomem_align and bs used. iykyk 🎀
hugepage-size=int — defines the size of a huge page. Must at least be equal to the system setting — go check /proc/meminfo and /sys/kernel/mm/hugepages/. Defaults to 2 or 4MiB depending on the platform. It should probably always be a multiple of megabytes, so using hugepage-size=Xm is the preferred way to set this, so you avoid setting a non-pow-2 bad value. Do it the cute way. 💅
lockmem=int — pin the specified amount of memory with mlock(2). Can be used to simulate a smaller amount of memory. The amount specified is per worker. 🫶

🌸 I/O size 🌸

size=int[%|z] — the total size of file I/O for each thread of this job. Fio will run until this many bytes have been transferred, unless the runtime gets altered by other means such as (1) runtime, (2) io_size, (3) number_ios, (4) gaps/holes while doing I/O's such as rw=read:16K, or (5) sequential I/O reaching end of the file, which is possible when percentage_random is less than 100. 😮

Fio will divide this size between the available files determined by options such as nrfiles, filename — unless filesize is specified by the job. If the result of that division happens to be 0, the size is set to the physical size of the given files or devices if they exist. If this option is not specified, fio will use the full size of the given files or devices. And if the files don't exist, size must be given (no vibes-based guessing here).

It's also possible to give size as a percentage between 1 and 100. If size=20% is given, fio will use 20% of the full size of the given files or devices. In ZBD mode, size can be given in units of number of zones using 'z'. Can be combined with offset to constrain the start and end range that I/O will be done within. 🦋

io_size=int[%|z], io_limit=int[%|z] — normally fio operates within the region set by size, which means the size option sets both the region and the size of I/O to be performed. Sometimes that's not what you want, and that's so valid. With this option it's possible to define just the amount of I/O that fio should do. For instance, if size is set to 20GiB and io_size is set to 5GiB, fio will perform I/O within the first 20GiB but exit when 5GiB have been done. The opposite is also possible — if size is set to 20GiB and io_size is set to 40GiB, then fio will do 40GiB of I/O within the 0..20GiB region. Literally the most flexible thing ever. ✨

Value can be set as a percentage: io_size=N%. In this case io_size multiplies the size= value. In ZBD mode, the value can also be set as number of zones using 'z'.

filesize=irange(int) — individual file sizes. May be a range, in which case fio will select sizes for files at random within the given range. If not given, each created file is the same size. This option overrides size in terms of file size, i.e. size becomes merely the default for io_size (and has no effect at all if io_size is set explicitly). 👑
file_append=bool — perform I/O after the end of the file. Normally fio operates within the size of a file; if this option is set, fio will append to the file instead. This has identical behavior to setting offset to the size of a file. This option is ignored on non-regular files. 🙈
fill_device=bool, fill_fs=bool — sets size to something really large and waits for ENOSPC (no space left on device) or EDQUOT (disk quota exceeded) as the terminating condition. Only makes sense with sequential write. For a read workload, the mount point will be filled first and then I/O started on the result. 🍓

🔥 I/O engine 🔥

ioengine=str — fio supports 2 kinds of performance measurement: I/O, and file/directory operation.

I/O engines define how the job issues I/O to the file. The following types are defined:

sync — basic read(2) or write(2) I/O. lseek(2) is used to position the I/O location. See fsync and fdatasync for syncing write I/Os.
psync — basic pread(2) or pwrite(2) I/O. Default on all supported operating systems except for Windows. 💅
vsync — basic readv(2) or writev(2) I/O. Will emulate queuing by coalescing adjacent I/Os into a single submission.
pvsync — basic preadv(2) or pwritev(2) I/O.
pvsync2 — basic preadv2(2) or pwritev2(2) I/O.
io_uring — fast Linux native asynchronous I/O. Supports async IO for both direct and buffered IO. This engine defines engine specific options. 🔥
io_uring_cmd — fast Linux native asynchronous I/O for passthrough commands. This engine defines engine specific options.
libaio — Linux native asynchronous I/O. Note that Linux may only support queued behavior with non-buffered I/O (set direct=1 or buffered=0). This engine defines engine specific options.
posixaio — POSIX asynchronous I/O using aio_read(3) and aio_write(3).
solarisaio — Solaris native asynchronous I/O.
windowsaio — Windows native asynchronous I/O. Default on Windows.
mmap — file is memory mapped with mmap(2) and data copied to/from using memcpy(3).
splice — splice(2) is used to transfer the data and vmsplice(2) to transfer data from user space to the kernel.
sg — SCSI generic sg v3 I/O. May either be synchronous using the SG_IO ioctl, or if the target is an sg character device we use read(2) and write(2) for asynchronous I/O. Requires the filename option to specify either block or character devices. This engine supports trim operations. The sg engine includes engine specific options. 🌈
libzbc — read, write, trim and ZBC/ZAC operations to a zoned block device using the libzbc library. The target can be either an SG character device or a block device file.
null — doesn't transfer any data, just pretends to. This is mainly used to exercise fio itself and for debugging/testing purposes. It's giving "fake it till you make it" 💀
net — transfer over the network to given host:port. Depending on the protocol used, the hostname, port, listen and filename options are used to specify what sort of connection to make, while the protocol option determines which protocol will be used. This engine defines engine specific options.
netsplice — like net, but uses splice(2) and vmsplice(2) to map data and send/receive. This engine defines engine specific options.
cpuio — doesn't transfer any data, but burns CPU cycles according to the cpuload, cpuchunks and cpumode options. A job never finishes unless there is at least one non-cpuio job. 😭

cpuload=85 will cause that job to do nothing but burn 85% of the CPU. In case of SMP machines, use numjobs=<nr_of_cpu> to get desired CPU usage, as the cpuload only loads a single CPU at the desired rate.

cpumode=qsort replaces the default noop instructions loop by a qsort algorithm to consume more energy. 🔥

rdma — the RDMA I/O engine supports both RDMA memory semantics (RDMA_WRITE/RDMA_READ) and channel semantics (Send/Recv) for the InfiniBand, RoCE and iWARP protocols. This engine defines engine specific options.
falloc — I/O engine that does regular fallocate to simulate data transfer as an fio ioengine:

DDIR_READ      does fallocate(,mode = FALLOC_FL_KEEP_SIZE,).
DIR_WRITE      does fallocate(,mode = 0).
DDIR_TRIM      does fallocate(,mode = FALLOC_FL_KEEP_SIZE|FALLOC_FL_PUNCH_HOLE).

ftruncate — I/O engine that sends ftruncate(2) operations in response to write (DDIR_WRITE) events. Each ftruncate issued sets the file's size to the current block offset. blocksize is ignored. 👀
e4defrag — I/O engine that does regular EXT4_IOC_MOVE_EXT ioctls to simulate defragment activity in request to DDIR_WRITE event.
rados — I/O engine supporting direct access to Ceph Reliable Autonomic Distributed Object Store (RADOS) via librados. This ioengine defines engine specific options. 🧚

`rbd`

Lowkey the coolest way to talk straight to Ceph Rados Block Devices (RBD) — it goes through librbd and skips the whole kernel rbd driver situation entirely. ✨ Heads up tho: this ioengine defines its own engine specific options, so go read those before you start vibing.

`http`

GET/PUT requests over HTTP(S) with libcurl, hitting a WebDAV or S3 endpoint. Straightforward, iconic, does what it says. 💅 This ioengine also defines engine specific options.

Real talk though — this engine only supports direct IO of iodepth=1, so if you want more parallelism you gotta scale it via numjobs. The blocksize defines the size of the objects to be created. And TRIM? It's translated to object deletion. 🗑️

`gfapi`

Uses the GlusterFS libgfapi sync interface to get direct access to GlusterFS volumes without having to go through FUSE. Iconic shortcut behaviour. This ioengine defines engine specific options.

`gfapi_async`

Same deal as gfapi but with the GlusterFS libgfapi async interface — direct access to GlusterFS volumes, no FUSE in the middle. 🦋 This ioengine defines engine specific options.

`libhdfs`

Read and write through Hadoop (HDFS). The filename option is what you use to specify the host,port of the hdfs name-node to connect to.

Now, the way this engine handles offsets is a little different — plot twist 🫢 — because in HDFS, files once created cannot be modified, so random writes are just not possible, periodt. To imitate that, the libhdfs engine expects a bunch of small files to be created over HDFS and will randomly pick a file from them based on the offset generated by the fio backend (peep the example job file to create such files, use the rw=write option). Also ngl, it may be necessary to set environment variables to work with HDFS/libhdfs properly. And each job uses its own connection to HDFS — no sharing, main character energy only.

`mtd`

Read, write and erase an MTD character device (e.g., /dev/mtd0). Discards are treated as erases — noted. Depending on the underlying device type, the I/O may have to go in a certain pattern, e.g., on NAND, writing sequentially to erase blocks and discarding before overwriting. The trimwrite mode works well for this constraint. 🎀

`dev-dax`

Read and write using device DAX to a persistent memory device (e.g., /dev/dax0.0) through the PMDK libpmem library. Simple, clean, slays.

`external`

This one's a prefix, not a full engine — it specifies loading an external I/O engine object file. Append the engine filename, e.g. ioengine=external:/tmp/foo.o to load ioengine foo.o in /tmp. The path can be either absolute or relative. 🌈 See engines/skeleton_external.c in the fio source for details of writing an external I/O engine.

`libpmem`

Read and write using mmap I/O to a file on a filesystem mounted with DAX on a persistent memory device, through the PMDK libpmem library. 💖

`ime_psync`

Synchronous read and write using DDN's Infinite Memory Engine (IME). This engine is very basic and issues calls to IME whenever an IO is queued. No thoughts, just calls. 💀

`ime_psyncv`

Synchronous read and write using DDN's Infinite Memory Engine (IME), but this one uses iovecs and will try to stack as much IOs as possible (if the IOs are "contiguous" and the IO depth is not exceeded) before issuing a call to IME. Efficiency queen. 👑

`ime_aio`

Asynchronous read and write using DDN's Infinite Memory Engine (IME). This engine will try to stack as much IOs as possible by creating requests for IME. FIO will then decide when to commit these requests — it's giving delegated authority. ✨

`libiscsi`

Read and write iscsi lun with libiscsi. Short, sweet, does the thing.

`nbd`

Synchronous read and write a Network Block Device (NBD). That's it, that's the vibe. 🦋

`libcufile`

I/O engine supporting libcufile synchronous access to nvidia-fs and a GPUDirect Storage-supported filesystem. This engine performs I/O without transferring buffers between user-space and the kernel, unless verify is set or cuda_io is posix. Also: iomem must not be cudamalloc. 🙈 This ioengine defines engine specific options.

`dfs`

I/O engine supporting asynchronous read and write operations to the DAOS File System (DFS) via libdfs. Async girlies rise. 💜

`nfs`

I/O engine supporting asynchronous read and write operations to NFS filesystems from userspace via libnfs. This is useful for achieving higher concurrency and thus throughput than is possible via kernel NFS. Lowkey a glow-up for your NFS numbers. 🔥

`exec`

Execute 3rd party tools. Could be used to perform monitoring during jobs runtime. Little side-quest engine, honestly kinda useful.

`xnvme`

I/O engine using the xNVMe C API, for NVMe devices. The xnvme engine provides flexibility to access GNU/Linux Kernel NVMe driver via libaio, IOCTLs, io_uring, the SPDK NVMe driver, or your own custom NVMe driver. The xnvme engine includes engine specific options. (See https://xnvme.io/.) 🌈

`libblkio`

Use the libblkio library (https://gitlab.com/libblkio/libblkio). The specific driver to use must be set using libblkio_driver. If mem/iomem is not specified, memory allocation is delegated to libblkio (and so is guaranteed to work with the selected driver). One libblkio instance is used per process, so all jobs setting option thread will share a single instance (with one queue per thread) and must specify compatible options. Note that some drivers don't allow several instances to access the same device or file simultaneously, but allow it for threads. 💅

🎀 File/directory operation engines 🎀

These engines define how the job operates a file or a directory — no actual data I/O, just file-system behaviour. The following types are defined:

filecreate — Simply create the files and do no I/O to them. You still need to set filesize so that all the accounting still occurs, but no actual I/O will be done other than creating the file. Example job file: filecreate-ioengine.fio.
filestat — Simply do stat() and do no I/O to the file. You need to set filesize and nrfiles, so that files will be created. This engine is to measure file lookup and meta data access. Example job file: filestat-ioengine.fio.
filedelete — Simply delete the files by unlink() and do no I/O to them. You need to set filesize and nrfiles, so that the files will be created. This engine is to measure file delete. Example job file: filedelete-ioengine.fio.
dircreate — Simply create the directories and do no I/O to them. You still need to set filesize so that all the accounting still occurs, but no actual I/O will be done other than creating the directories. Example job file: dircreate-ioengine.fio.
dirstat — Simply do stat() and do no I/O to the directories. You need to set filesize and nrfiles, so that directories will be created. This engine is to measure directory lookup and meta data access. Example job file: dirstat-ioengine.fio.
dirdelete — Simply delete the directories by rmdir() and do no I/O to them. You need to set filesize and nrfiles, so that the directories will be created. This engine is to measure directory delete.

⚠️ For file and directory operation engines, there is no I/O throughput, so the statistics data in report have different meanings. The meaningful output indexes are: iops and clat. bw is meaningless. Refer to section: "Interpreting the output" for more details.

🔥 I/O engine specific parameters 🔥

On top of everything else, there are some parameters which are only valid when a specific ioengine is in use. No cap. These are used identically to normal parameters, with the caveat that when used on the command line, they must come after the ioengine that defines them is selected. (Order matters, bestie — don't come for me, that's just how it is. 😭)

(io_uring,libaio)cmdprio_percentage=int[,int] — Set the percentage of I/O that will be issued with the highest priority. Default: 0. A single value applies to reads and writes. Comma-separated values may be specified for reads and writes. For this option to be effective, NCQ priority must be supported and enabled, and the direct=1 option must be used. fio must also be run as the root user. 👑 Unlike slat/clat/lat stats, which can be tracked and reported independently, per priority stats only track and report a single type of latency. By default, completion latency (clat) will be reported, if lat_percentiles is set, total latency (lat) will be reported.
(io_uring,libaio)cmdprio_class=int[,int] — Set the I/O priority class to use for I/Os that must be issued with a priority when cmdprio_percentage or cmdprio_bssplit is set. If not specified when cmdprio_percentage or cmdprio_bssplit is set, this defaults to the highest priority class. A single value applies to reads and writes. Comma-separated values may be specified for reads and writes. See man ionice(1). See also the prioclass option.
(io_uring,libaio)cmdprio_hint=int[,int] — Set the I/O priority hint to use for I/Os that must be issued with a priority when cmdprio_percentage or cmdprio_bssplit is set. If not specified when cmdprio_percentage or cmdprio_bssplit is set, this defaults to 0 (no hint). A single value applies to reads and writes. Comma-separated values may be specified for reads and writes. See also the priohint option. ✨
(io_uring,libaio)cmdprio=int[,int] — okay so this is the actual priority value for I/Os that gotta go out with a priority, and it only kicks in when cmdprio_percentage or cmdprio_bssplit is set. If you don't set it while one of those is active, it just defaults to 0, no drama. Linux caps you to a positive value between 0 and 7, and plot twist: 0 is the highest priority, not the lowest, iykyk. One single value applies to both reads and writes, or you can drop comma-separated values for reads and writes separately. Go peep ionice(1) for the lore. On other operating systems the meaning of priority may differ so check their manpages, bestie. Also see the prio option. 🎀
(io_uring,libaio)cmdprio_bssplit=str[,str] — this one's for when you want finer control over I/O priority: it lets you say what percentage of IOs must have a priority set depending on the block size of the IO. It's only useful when you use it with the bssplit option, like, multiple different block sizes for reads and writes. 💅

First accepted format is literally the same as the format of the bssplit option:

cmdprio_bssplit=blocksize/percentage:blocksize/percentage

With this one, every entry uses the priority class, priority hint and priority level defined by the cmdprio_class, cmdprio and cmdprio_hint options respectively.

Second accepted format, and this is the glow-up:

cmdprio_bssplit=blocksize/percentage/class/level:blocksize/percentage/class/level

Here the priority class and priority level live inside each entry. Compared to the first format, this one doesn't force all the entries to share the same priority class and priority level. freedom. 🦋

Third accepted format, ok she's fancy:

cmdprio_bssplit=blocksize/percentage/class/level/hint:...

It's an extension of the second format that also lets you specify a priority hint.

For all the formats, only the read and write data directions are supported, and values for trim IOs get ignored. This option is mutually exclusive with the cmdprio_percentage option — pick a lane, babe.

(io_uring,io_uring_cmd)fixedbufs — so if fio is asked to do direct IO, Linux maps pages for every IO call and releases them when the IO is done. Set this and the pages get pre-mapped before IO even starts. No more map-and-release for every single IO, which is more efficient and lowkey reduces the IO latency too. obsessed. ✨
(io_uring,io_uring_cmd)nonvectored=int — with this, fio uses non-vectored read/write commands, where the address must contain the address directly. Default is -1.
(io_uring,io_uring_cmd)force_async — normally io_uring tries to issue an sqe as non-blocking first, and if that fails it executes it in an async manner. Set this to N and every N request, fio asks the sqe to be issued async. Default is 0.
(io_uring,io_uring_cmd,xnvme)hipri — set this and fio will attempt to use polled IO completions. Normal IO completions generate interrupts to signal completion; polled completions do not, so they require active reaping by the application. The payoff is more efficient IO for high IOPS scenarios and lower latencies for low queue depth IO. 💅
(io_uring,io_uring_cmd)registerfiles — with this, fio registers the set of files being used with the kernel. That dodges the overhead of managing file counts in the kernel, making submission and completion more lightweight. Required for the sqthread_poll option below, so don't skip it.
(io_uring,io_uring_cmd,xnvme)sqthread_poll — normally fio submits IO by issuing a system call to notify the kernel of available items in the SQ ring. Set this and the submitting gets done by a polling thread in the kernel. That frees up cycles for fio, at the cost of using more CPU in the system. And because submission is just the time it takes to fill in the sqe entries plus any syscall needed to wake up the idle kernel thread, fio will not report submission latencies. know that. 🫶
(io_uring,io_uring_cmd)sqthread_poll_cpu=int — when sqthread_poll is set, this is how you define which CPU the polling thread should use.
(io_uring_cmd)cmd_type=str — specifies the type of uring passthrough command to use. The supported value is nvme, and the default is nvme. (yes bestie, same word twice, that's the whole bit 💀)
(libaio)userspace_reap — normally with the libaio engine, fio uses the io_getevents(3) system call to reap newly returned events. Turn this flag on and the AIO ring is read directly from user-space to reap events. Heads up: the reaping mode is only enabled when polling for a minimum of 0 events (e.g. when iodepth_batch_complete=0). 🔥
(pvsync2)hipri — sets RWF_HIPRI on I/O, telling the kernel it's higher priority than normal.
(pvsync2)hipri_percentage — when hipri is set, this determines the probability of a pvsync2 I/O being high priority. Default is 100%.
(pvsync2,libaio,io_uring,io_uring_cmd)nowait=bool — by default, if a request can't be executed immediately (resource starvation, waiting on locks, the usual), it gets queued and the initiating process is blocked until the required resource frees up. This option sets the RWF_NOWAIT flag (supported from the 4.14 Linux kernel) and the call returns instantly with EAGAIN or a partial result instead of waiting. 🙈

It's useful to also use ignore_error=EAGAIN when you use this. Note: glibc 2.27 and 2.28 have a bug in the syscall wrappers preadv2, pwritev2 — they return EOPNOTSUP instead of EAGAIN. 😭

For cached I/O, using this usually means a request operates only with cached data. Currently the RWF_NOWAIT flag is not supported for cached write. For direct I/O, requests only succeed if cache invalidation isn't required, file blocks are fully allocated, and the disk request could be issued immediately. 💅

(pvsync2,io_uring)uncached=int — this one does buffered IO without retaining data in the page cache after the operation completes. Reads behave like a normal buffered read but the pages get evicted immediately after data is copied to userspace. Writes behave like buffered writes but a writeback is initiated before the syscall returns, and pages are evicted once the writeback completes. It sets the RWF_UNCACHED flag (supported from the 6.14 Linux kernel) on a per-IO basis. 🦋
(pvsync2,libaio,io_uring)atomic=bool — this means writes are issued with torn-write protection, so on a power fail or kernel crash, all or none of the data from the write gets stored, but never a mix of old and new data. Torn-write protection is also known as atomic writes and she's iconic. ✨

It sets the RWF_ATOMIC flag (supported from the 6.11 Linux kernel) on a per-IO basis.

Writes with RWF_ATOMIC set will be rejected by the kernel when the file does not support torn-write protection. To learn a file's torn-write limits, issue statx with STATX_WRITE_ATOMIC. 👑

(io_uring_cmd,xnvme)fdp=bool — enable Flexible Data Placement mode for write commands.
(io_uring_cmd,xnvme)dataplacement=str — specifies the data placement directive type to use for write commands. Supported types, bestie:
- none — do not use a data placement directive. This is the default.
- fdp — use Flexible Data placement directives for write commands. This is equivalent to specifying fdp=1.
- streams — use Streams directives for write commands.
(io_uring_cmd,xnvme)plid_select=str,fdp_pli_select=str — defines how fio decides which placement ID to use next. The types:
- random — choose a placement ID at random (uniform).
- roundrobin — round robin over available placement IDs. This is the default.
- scheme — choose a placement ID (index) based on the scheme file defined by the option dp_scheme.

The available placement ID (indices) are defined by plids or fdp_pli option, except in the case of scheme.

(io_uring_cmd,xnvme)plids=str,fdp_pli=str — pick which Placement ID Indices (FDP) or Placement IDs (streams) this job is allowed to use for writes. It accepts a comma-separated list of values or ranges (e.g. 1,2-4,5,6-8). 🌈

For FDP, by default the job cycles through all available Placement IDs, so use this to be selective. The values here are array indices for the list of placement IDs returned by the nvme-cli command nvme fdp status. If you want fio to use FDP placement identifiers only at indices 0, 2 and 5, set plids=0,2,5. 🎯

For streams, this should be a list of Stream IDs.

(io_uring_cmd,xnvme)dp_scheme=str — defines which placement ID (index) gets selected based on offset(LBA) range. The file should contain one or more scheme entries in this format:

0, 10737418240, 0
10737418240, 21474836480, 1
21474836480, 32212254720, 2
...

Each line, a scheme entry, contains start offset, end offset, and placement ID (index) separated by comma (,). If the write offset is within the range of a certain scheme entry (start offset ≤ offset < end offset), the corresponding placement ID (index) gets selected. If the write offset belongs to multiple scheme entries, the first matched scheme entry is applied. If the offset isn't within any range of a scheme entry, the dspec field is set to 0, default RUH.

Caution: in case of multiple devices in a job, all devices of the job will be affected by the scheme. If this option is specified, the option plids or fdp_pli will be ignored. 💀

okay so these next ones are all per-engine knobs, and honestly they're giving "very specific but very important" 💅. every single one matters, no skipping, let's go.

✨ the nvme / io_uring_cmd girls ✨

(io_uring_cmd,xnvme)md_per_io_size=int — ok so this is the size in bytes for a separate metadata buffer, per IO. heads up: for io_uring_cmd these buffers get allocated using malloc no matter what you set for iomem. Default: 0.
(io_uring_cmd,xnvme)pi_act=int — the action to take when the nvme namespace is formatted with protection information. break it down:
- if this is 1 and the namespace is formatted with metadata size equal to protection information size, fio won't use a separate metadata buffer or extended logical block.
- if this is 1 and the namespace is formatted with metadata size greater than protection information size, fio will not generate or verify the protection information portion of metadata for the write or read case respectively.
- if this is 0, fio generates protection information for the write case and verifies it for the read case.

Default: 1.

also for 16 bit CRC generation, fio will use isa-l if it's available, otherwise it uses the default slower generator (see: https://github.com/intel/isa-l) 🧚

(io_uring_cmd,xnvme)pi_chk=str[,str][,str] — controls the protection information check. you can pass one or more of these values. Default: none.
- GUARD — enables protection information checking of the guard field.
- REFTAG — enables protection information checking of the logical block reference tag field.
- APPTAG — enables protection information checking of the application tag field.
(io_uring_cmd,xnvme)apptag=int — specifies the logical block application tag value, if the namespace is formatted to use end to end protection information. Default: 0x1234.
(io_uring_cmd,xnvme)apptag_mask=int — same vibe, but the logical block application tag mask value, if the namespace is formatted to use end to end protection information. Default: 0xffff.
(io_uring_cmd)num_range=int — for the trim command, this is the number of ranges to trim per I/O request. the number of logical blocks per range is decided by the bs option, which should be a multiple of the logical block size. this cannot be used with read or write, periodt. also note: if you set this option > 1, log_offset will not be able to log all the offsets 😭. Default: 1.

🎀 cpuio (aka the "make my cpu sweat" engine) 🎀

(cpuio)cpuload=int — attempt to use the specified percentage of CPU cycles. this one is mandatory when you're using the cpuio I/O engine, so don't forget it bestie.
(cpuio)cpuchunks=int — splits the load into cycles of the given time. in microseconds.
(cpuio)cpumode=str — specifies how to stress the CPU. it takes these two values:
- noop — this is the default, and it directs the CPU to execute noop instructions.
- qsort — replaces the default noop instructions with a qsort algorithm to consume more energy. 🔥
(cpuio)exit_on_io_done=bool — detect when the I/O threads are done, then exit.

🦋 hdfs moment 🦋

(libhdfs)namenode=str — the hostname or IP address of a HDFS cluster namenode to contact.
(libhdfs)port=int — the listening port of the HFDS cluster namenode. (that's "HFDS", not a typo of mine, i'm literally quoting it 💀)

🌸 network stuff: netsplice, net, rdma 🌸

(netsplice,net)port=int — the TCP or UDP port to bind to or connect to. if you use this with numjobs to spawn multiple instances of the same job type, then this is the starting port number, since fio will use a range of ports.
(rdma)port=int — the port to use for RDMA-CM communication. this should be the same value on the client and the server side, no cap.
(netsplice,net,rdma)hostname=str — the hostname or IP address to use for TCP, UDP or RDMA-CM based I/O. if the job is a TCP listener or UDP reader, the hostname isn't used and must be omitted — unless it's a valid UDP multicast address, in which case it's allowed.
(netsplice,net)interface=str — the IP address of the network interface used to send or receive UDP multicast.
(netsplice,net)ttl=int — time-to-live value for outgoing UDP multicast packets. Default: 1.
(netsplice,net)nodelay=bool — set TCP_NODELAY on TCP connections.
(netsplice,net)protocol=str, proto=str — the network protocol to use. accepted values are:
- tcp — transmission control protocol.
- tcpv6 — transmission control protocol V6.
- udp — user datagram protocol.
- udpv6 — user datagram protocol V6.
- unix — UNIX domain socket.
- vsock — VSOCK protocol.

when the protocol is tcp, udp or vsock, the port must also be given, as well as the hostname if the job is a TCP or VSOCK listener or UDP reader. for unix sockets, the normal filename option should be used and the port is invalid. when the protocol is vsock, the hostname is the CID of the remote VM. iykyk 🧚

(netsplice,net)listen — for TCP network connections, this tells fio to listen for incoming connections instead of initiating an outgoing connection. the hostname must be omitted if this option is used.
(netsplice,net)pingpong — normally a network writer just keeps writing data and a network reader just consumes packages. but if pingpong=1 is set, a writer sends its normal payload to the reader, then waits for the reader to send the same payload back. this lets fio measure network latencies 🫶. the submission and completion latencies then measure local time spent sending or receiving, and the completion latency measures how long it took for the other end to receive and send back. for UDP multicast traffic, pingpong=1 should only be set for a single reader when multiple readers are listening to the same address.
(netsplice,net)window_size=int — set the desired socket buffer size for the connection.
(netsplice,net)mss=int — set the TCP maximum segment size (TCP_MAXSEG).

💜 e4defrag 💜

(e4defrag)donorname=str — the file will be used as a block donor (swap extents between files).
(e4defrag)inplace=int — configure the donor file blocks allocation strategy:
- 0 — default. preallocate donor's file on init.
- 1 — allocate space immediately inside the defragment event, and free right after the event.

🌈 ceph: rbd + rados 🌈

(rbd,rados)clustername=str — specifies the name of the Ceph cluster.
(rbd)rbdname=str — specifies the name of the RBD.
(rbd,rados)pool=str — specifies the name of the Ceph pool containing RBD or RADOS data.
(rbd,rados)clientname=str — specifies the username (without the client. prefix) used to access the Ceph cluster. if the clustername is specified, the clientname shall be the full type.id string. if no type. prefix is given, fio will add client. by default.
(rados)conf=str — specifies the configuration path of the ceph cluster, so the conf file doesn't have to be /etc/ceph/ceph.conf.
(rbd,rados)busy_poll=bool — poll store instead of waiting for completion. usually this gives better throughput at the cost of higher (up to 100%) CPU utilization 😭.
(rados)touch_objects=bool — during initialization, touch (create if they don't exist) all objects (files). touching all objects affects ceph caches and likely impacts test results. enabled by default.
(rbd)rbd_encryption_format=str — specifies the encryption format of the RBD image. supported values are luks1 and luks2. if you set this, rbd_encryption_passphrase must also be specified. note that the image must have been previously formatted using rbd encryption format <image> — the fio rbd engine will only attempt to load the encryption context, not format the image. the RBD encryption feature is disabled by default. support for this feature requires librbd version 16.2 (Ceph Pacific) or later.
(rbd)rbd_encryption_passphrase=str — the passphrase used to unlock the encrypted RBD image. required if rbd_encryption_format is set. 🔐

⭐ http (webdav / swift / s3) ⭐

(http)http_host=str — hostname to connect to. HTTP port 80 is used automatically when the value of the https parameter is off, and HTTPS port 443 if it's on. a virtual-hosted-style S3 hostname starts with a bucket name, while a path-style S3 hostname does not. default is localhost.
(http)http_user=str — username for HTTP authentication.
(http)http_pass=str — password for HTTP authentication.
(http)https=str — whether to use HTTPS instead of plain HTTP. on enables HTTPS; insecure will enable HTTPS, but disable SSL peer verification (use with caution! 🙈). default is off.
(http)http_mode=str — which HTTP access mode to use: webdav, swift, or s3. default is webdav.
(http)http_s3_region=str — the S3 region/zone to include in the request. default is us-east-1.
(http)http_s3_key=str — the S3 secret key.
(http)http_s3_keyid=str — the S3 key/access id.
(http)http_s3_security_token=str — the S3 security token.
(http)http_s3_sse_customer_key=str — the encryption customer key in SSE server side.
(http)http_s3_sse_customer_algorithm=str — the encryption customer algorithm in SSE server side. default is AES256.
(http)http_s3_storage_class=str — which storage class to access. user-customizable settings. default is STANDARD.
(http)http_swift_auth_token=str — the Swift auth token. see the example configuration file on how to retrieve this.
(http)http_verbose=int — enable verbose requests from libcurl. useful for debugging. 1 turns on verbose logging from libcurl, 2 additionally enables HTTP IO tracing. default is 0.
(http)http_object_mode=str — how to structure objects for HTTP IO: block or range. default is block.
- block — one object is created for every block. the HTTP engine treats blocksize as the size of the object to read or write, and appends the block start/end offsets to the filename to create the target object path. reads and writes operate on whole objects at a time.
range — one object per file, bestie 🎀 the object path is literally just the filename for both read and write I/O. For read requests, the blocksize and offset get used to set the "Range" header on read requests so you can issue partial reads of the object. For write requests, blocksize sets the size of the object, same as in block mode ✨
(mtd)skip_bad=bool — skip operations against known bad blocks 🙈
(libhdfs)hdfsdirectory — libhdfs will create the chunk in this HDFS directory 🦋
(libhdfs)chunk_size — the size of the chunk to use for each file 💅

🌈 RDMA things

(rdma)verb=str — the RDMA verb to use on this side of the RDMA ioengine connection. Valid values are write, read, send and recv. These correspond to the equivalent RDMA verbs (like write = rdma_write etc.). Note that this only needs to be specified on the client side of the connection 👑 see the examples folder!
(rdma)bindname=str — the name used to bind the local RDMA-CM connection to a local RDMA device. Could be a hostname or an IPv4 or IPv6 address. On the server side this gets passed into the rdma_bind_addr() function, and on the client side it's used in the rdma_resolve_add() function. Super useful when multiple paths exist between client and server, or in certain loopback configurations 💫

(filestat)stat_type=str — specify which stat system call type to measure lookup/getattr performance. Default is stat for stat(2) ⭐

🔥 sg / io_uring_cmd things

(sg)hipri — if this option is set, fio attempts to use polled IO completions, which has a similar effect to (io_uring)hipri. Only SCSI READ and WRITE commands get the SGV4_FLAG_HIPRI set — not UNMAP (trim) and not VERIFY 🫶 older versions of the Linux sg driver that don't support hipri will just ignore this flag and do normal IO. The Linux SCSI Low Level Driver (LLD) that "owns" the device also needs to support hipri (aka iopoll and mq_poll). The MegaRAID driver is an example of a SCSI LLD. Default: clear (0), which does normal (interrupted based) IO 💀
(sg,io_uring_cmd)readfua=bool — with readfua set to 1, read operations include the force unit access (fua) flag. Default: 0 🌸
(sg,io_uring_cmd)writefua=bool — with writefua set to 1, write operations include the force unit access (fua) flag. Default: 0 🌸
(io_uring_cmd)write_mode=str — specifies the type of write operation. Defaults to 'write' ✨
- write — use Write commands for write operations
- uncor — use Write Uncorrectable commands for write operations
- zeroes — use Write Zeroes commands for write operations
- verify — use Verify commands for write operations
(io_uring_cmd)verify_mode=str — specifies the type of command used in the verification phase. Defaults to 'read' 💜
- read — use Read commands for data verification
- compare — use Compare commands for data verification. This option is only valid with specific pattern(s), which means it must be given with verify=pattern and verify_pattern=<pattern> 👀
(sg)sg_write_mode=str — specify the type of write commands to issue. This option can take multiple values:
- write (default) — write opcodes are issued as usual
- write_and_verify — issue WRITE AND VERIFY commands. The BYTCHK bit is set to 00b. This directs the device to carry out a medium verification with no data comparison for the data that was written. The writefua option is ignored with this selection 🙈
- verify — this option is deprecated. Use write_and_verify instead ⚠️
- write_same — issue WRITE SAME commands. This transfers a single block to the device and writes that same block of data to a contiguous sequence of LBAs beginning at the specified offset. fio's blocksize parameter specifies the amount of data written with each command. However, the amount of data actually transferred to the device equals the device's block (sector) size. For a device with 512 byte sectors, blocksize=8k will write 16 sectors with each command. fio will still generate 8k of data for each command, but only the first 512 bytes get used and transferred to the device. The writefua option is ignored with this selection 💀
- same — this option is deprecated. Use write_same instead ⚠️
- write_same_ndob — issue WRITE SAME(16) commands as above but with the No Data Output Buffer (NDOB) bit set. No data will be transferred to the device with this bit set. Data written will be a pre-determined pattern such as all zeroes ✨
- write_stream — issue WRITE STREAM(16) commands. Use the stream_id option to specify the stream identifier 🌊
- verify_bytchk_00 — issue VERIFY commands with BYTCHK set to 00. This directs the device to carry out a medium verification with no data comparison 👑
- verify_bytchk_01 — issue VERIFY commands with BYTCHK set to 01. This directs the device to compare the data on the device with the data transferred to the device 🔍
- verify_bytchk_11 — issue VERIFY commands with BYTCHK set to 11. This transfers a single block to the device and compares the contents of that block with the data on the device beginning at the specified offset. fio's blocksize parameter specifies the total amount of data compared with this command. However, only one block (sector) worth of data is transferred to the device. Similar to the WRITE SAME command, except the data is compared instead of written 🫶
(sg)stream_id=int — set the stream identifier for WRITE STREAM commands. If this is set to 0 (which is not a valid stream identifier) fio will open a stream and then close it when done. Default is 0 💅

(nbd)uri=str — specify the NBD URI of the server to test. The string is a standard NBD URI (see https://github.com/NetworkBlockDevice/nbd/tree/master/doc). Example URIs:

nbd://localhost:10809

nbd+unix:///?socket=/tmp/socket

nbds://tlshost/exportname

🧚 libcufile things

(libcufile)gpu_dev_ids=str — specify the GPU IDs to use with CUDA. This is a colon-separated list of int. GPUs are assigned to workers roundrobin. Default is 0 ⭐
(libcufile)cuda_io=str — specify the type of I/O to use with CUDA. This option takes the following values:
- cufile (default) — use libcufile and nvidia-fs. This option performs I/O directly between a GPUDirect Storage filesystem and GPU buffers, avoiding use of a bounce buffer. If verify is set, cudaMemcpy is used to copy verification data between RAM and GPU(s). Verification data is copied from RAM to GPU before a write, and from GPU to RAM after a read. direct must be 1 🔥
- posix — use POSIX to perform I/O with a RAM buffer, and use cudaMemcpy to transfer data between RAM and the GPU(s). Data is copied from GPU to RAM before a write, and copied from RAM to GPU after a read. verify does not affect the use of cudaMemcpy 🦋

🌸 DAOS / dfs things

(dfs)pool — specify the label or UUID of the DAOS pool to connect to 💜
(dfs)cont — specify the label or UUID of the DAOS container to open 🎀
(dfs)chunk_size — specify a different chunk size (in bytes) for the dfs file. Uses the DAOS container's chunk size by default ✨
(dfs)object_class — specify a different object class for the dfs file. Uses the DAOS container's object class by default ✨

(nfs)nfs_url — URL in libnfs format, eg nfs://<server|ipv4|ipv6>/path[?arg=val[&arg=val]*] — refer to the libnfs README for more details 🫶

🎀 exec things

(exec)program=str — specify the program to execute. Heads up: the program will receive a SIGTERM when the job is reaching the time limit. A SIGKILL is sent once the job is over. The delay between the two signals is defined by the grace_time option 💀
(exec)arguments=str — specify arguments to pass to program. Some special variables can be expanded to pass fio's job details to the program:
- %r — replaced by the duration of the job in seconds
- %n — replaced by the name of the job
(exec)grace_time=int — defines the time between the SIGTERM and SIGKILL signals. Default is 1 second ⏱️
(exec)std_redirect=bool — if set, stdout and stderr streams are redirected to files named from the job name. Default is true 🌈

🦋 xnvme things

(xnvme)xnvme_async=str — select the xnvme async command interface. This can take these values:
- emu — this is the default, and is used to emulate asynchronous I/O by using a single thread to create a queue pair on top of a synchronous I/O interface using the NVMe driver IOCTL 👑
- thrpool — emulate an asynchronous I/O interface with a pool of userspace threads on top of a synchronous I/O interface using the NVMe driver IOCTL. By default four threads are used 🧵
io_uring — the Linux native asynchronous I/O interface, and it supports both direct and buffered I/O. versatile icon 💅
libaio — uses Linux aio for Asynchronous I/O.
posix — uses the posix asynchronous I/O interface to perform one or more I/O operations asynchronously.
vfio — uses the user-space VFIO-based backend, implemented using libvfn instead of SPDK.
nil — does not transfer any data, it just pretends to 🫥 mainly used for introspective performance evaluation. fake it till you make it fr.

(xnvme)xnvme_sync=str — picks the xnvme synchronous command interface. the values:
- nvme — this is the default and uses Linux NVMe Driver ioctl() for synchronous I/O.
- psync — this supports regular as well as vectored pread() and pwrite() commands.
- block — this is the same as psync except that it also supports zone management commands using Linux block layer IOCTLs. bougie upgrade ✨
(xnvme)xnvme_admin=str — picks the xnvme admin command interface. the values:
- nvme — this is the default and uses Linux NVMe Driver ioctl() for admin commands.
- block — use Linux Block Layer ioctl() and sysfs for admin commands.
(xnvme)xnvme_dev_nsid=int — the xnvme namespace identifier for the userspace NVMe driver SPDK or vfio.
(xnvme)xnvme_dev_subnqn=str — sets the subsystem NQN for fabrics. this is for xNVMe to utilize a fabrics target with multiple systems 🧵
(xnvme)xnvme_mem=str — picks the xnvme memory backend. the values:
- posix — this is the default posix memory backend for linux NVMe driver.
- hugepage — use hugepages instead of the existing posix memory backend. the memory backend uses hugetlbfs. this requires users to allocate hugepages, mount hugetlbfs and set an environment variable for XNVME_HUGETLB_PATH.
- spdk — uses SPDK's memory allocator.
- vfio — uses libvfn's memory allocator. this also specifies the use of the libvfn backend instead of SPDK.
(xnvme)xnvme_iovec — if this option is set, xnvme will use vectored read/write commands 🎀

(libblkio)libblkio_driver=str — the libblkio driver to use. different drivers access devices through different underlying interfaces. the available drivers depend on the libblkio version in use and are listed at https://libblkio.gitlab.io/libblkio/blkio.html#drivers
(libblkio)libblkio_path=str — sets the value of the driver-specific "path" property before connecting the libblkio instance, which identifies the target device or file on which to perform I/O. its exact semantics are driver-dependent and not all drivers may support it; see https://libblkio.gitlab.io/libblkio/blkio.html#drivers
(libblkio)libblkio_pre_connect_props=str — a colon-separated list of additional libblkio properties to be set after creating but before connecting the libblkio instance. each property must have the format <name>=<value>. colons can be escaped as \:. these are set after the engine sets any other properties, so those can be overridden 😌 available properties depend on the libblkio version in use and are listed at https://libblkio.gitlab.io/libblkio/blkio.html#properties
(libblkio)libblkio_num_entries=int — sets the value of the driver-specific "num-entries" property before starting the libblkio instance. its exact semantics are driver-dependent and not all drivers may support it; see https://libblkio.gitlab.io/libblkio/blkio.html#drivers
(libblkio)libblkio_queue_size=int — sets the value of the driver-specific "queue-size" property before starting the libblkio instance. its exact semantics are driver-dependent and not all drivers may support it; see https://libblkio.gitlab.io/libblkio/blkio.html#drivers
(libblkio)libblkio_pre_start_props=str — a colon-separated list of additional libblkio properties to be set after connecting but before starting the libblkio instance. each property must have the format <name>=<value>. colons can be escaped as \:. these are set after the engine sets any other properties, so those can be overridden 🫶 available properties depend on the libblkio version in use and are listed at https://libblkio.gitlab.io/libblkio/blkio.html#properties
(libblkio)hipri — use poll queues. heads up bestie, this is incompatible with libblkio_wait_mode=eventfd and libblkio_force_enable_completion_eventfd 💀
(libblkio)libblkio_vectored — submit vectored read and write requests.
(libblkio)libblkio_write_zeroes_on_trim — submit trims as "write zeroes" requests instead of discard requests.
(libblkio)libblkio_wait_mode=str — how to wait for completions:
- block (default) — use a blocking call to blkioq_do_io().
- eventfd — use a blocking call to read() on the completion eventfd.
- loop — use a busy loop with a non-blocking call to blkioq_do_io().
(libblkio)libblkio_force_enable_completion_eventfd — enable the queue's completion eventfd even when unused. this may impact performance 🐌 the default is to enable it only if libblkio_wait_mode=eventfd.
(windowsaio)no_completion_thread — avoid using a separate thread for completion polling.

🌊 I/O depth 🌊

iodepth=int — number of I/O units to keep in flight against the file. note that increasing iodepth beyond 1 will not affect synchronous ioengines (except for small degrees when verify_async is in use). even async engines may impose OS restrictions causing the desired depth not to be achieved. this may happen on Linux when using libaio and not setting direct=1, since buffered I/O is not async on that OS 😭 keep an eye on the I/O depth distribution in the fio output to verify that the achieved depth is as expected. Default: 1.
iodepth_batch_submit=int, iodepth_batch=int — this defines how many pieces of I/O to submit at once. it defaults to 1, which means that we submit each I/O as soon as it is available, but can be raised to submit bigger batches of I/O at the time. if it is set to 0 the iodepth value will be used.
iodepth_batch_complete_min=int, iodepth_batch_complete=int — this defines how many pieces of I/O to retrieve at once. it defaults to 1, which means that we'll ask for a minimum of 1 I/O in the retrieval process from the kernel. the I/O retrieval will go on until we hit the limit set by iodepth_low. if this variable is set to 0, then fio will always check for completed events before queuing more I/O. this helps reduce I/O latency, at the cost of more retrieval system calls ✨
iodepth_batch_complete_max=int — this defines the maximum pieces of I/O to retrieve at once. this variable should be used along with iodepth_batch_complete_min=int, specifying the range of min and max amount of I/O which should be retrieved. by default it is equal to the iodepth_batch_complete_min value. Example #1:

iodepth_batch_complete_min=1
iodepth_batch_complete_max=<iodepth>

which means that we will retrieve at least 1 I/O and up to the whole submitted queue depth. if none of the I/O has been completed yet, we will wait.

Example #2:

iodepth_batch_complete_min=0
iodepth_batch_complete_max=<iodepth>

which means that we can retrieve up to the whole submitted queue depth, but if none of the I/O has been completed yet, we will NOT wait and immediately exit the system call. in this example we simply do polling 👀

iodepth_low=int — the low water mark indicating when to start filling the queue again. defaults to the same as iodepth, meaning that fio will attempt to keep the queue full at all times. if iodepth is set to e.g. 16 and iodepth_low is set to 4, then after fio has filled the queue of 16 requests, it will let the depth drain down to 4 before starting to fill it again. it's giving water bottle refill energy 💧
serialize_overlap=bool — serialize in-flight I/Os that might otherwise cause or suffer from data races. when two or more I/Os are submitted simultaneously, there is no guarantee that the I/Os will be processed or completed in the submitted order 😬 further, if two or more of those I/Os are writes, any overlapping region between them can become indeterminate/undefined on certain storage. these issues can cause verification to fail erratically when at least one of the racing I/Os is changing data and the overlapping region has a non-zero size. setting serialize_overlap tells fio to avoid provoking this behavior by explicitly serializing in-flight I/Os that have a non-zero overlap. note that setting this option can reduce both performance and the iodepth achieved.

this option only applies to I/Os issued for a single job except when it is enabled along with io_submit_mode=offload. in offload mode, fio will check for overlap among all I/Os submitted by offload jobs with serialize_overlap enabled.

Default: false.

io_submit_mode=str — this option controls how fio submits the I/O to the I/O engine. the default is inline, which means that the fio job threads submit and reap I/O directly. if set to offload, the job threads will offload I/O submission to a dedicated pool of I/O threads 🔥 this requires some coordination and thus has a bit of extra overhead, especially for lower queue depth I/O where it can increase latencies. the benefit is that fio can manage submission rates independently of the device completion rates. this avoids skewed latency reporting if I/O gets backed up on the device side (the coordinated omission problem).

note: this option cannot reliably be used with async IO engines 🙈

✨ I/O rate ✨

thinkcycles=int — makes your job literally stall for this many cycles after an I/O finishes, before it's allowed to fire the next one 🌸. Use it to pretend your app is doing actual ~processing~ on the data. Heads up tho: this does not get factored into the time you wait out for thinktime. Also it might just do nothing on some platforms 😭 — you can vibe-check that by setting a high enough amount of thinkcycles and seeing if anything happens.
thinktime=time — stall the job for this period of time after an I/O completes, before issuing the next one. Same deal: it's for simulating an application doing processing work. When you leave the unit off, the value is interpreted in microseconds 🫶. See thinktime_blocks, thinktime_iotime and thinktime_spin.
thinktime_spin=time — only valid if thinktime is set 💅. It makes fio pretend to burn CPU on the data it just received, before falling back to sleeping for the rest of the period specified by thinktime. When the unit is omitted, the value is interpreted in microseconds.
thinktime_blocks=int — only valid if thinktime is set, bestie. It controls how many blocks get issued before waiting thinktime usecs. If it's not set, it defaults to 1, which makes fio wait thinktime usecs after every block. And like… that effectively makes any queue depth setting redundant, because no more than 1 I/O will be queued before we have to complete it and do our thinktime. In other words, this setting effectively caps the queue depth if the latter is larger 😭.
thinktime_blocks_type=str — only valid if thinktime is set, and it controls how thinktime_blocks triggers 👑. The default is complete, which triggers thinktime when fio completes thinktime_blocks blocks. If you set it to issue, then the trigger happens at the issue side instead.
thinktime_iotime=time — only valid if thinktime is set. This one controls the thinktime interval by time ⏰. The thinktime stall gets repeated after IOs are executed for thinktime_iotime. Like, --thinktime_iotime=9s --thinktime=1s repeats a 10-second cycle — IOs for 9 seconds, then stall for 1 second, iconic. When the unit is omitted, thinktime_iotime is interpreted as a number of seconds. And if you use it together with thinktime_blocks, the thinktime stall is repeated after thinktime_iotime or after thinktime_blocks IOs — whichever happens first 🔥.
rate=int[,int][,int] — caps the bandwidth this job is allowed to use. The number is in bytes/sec, and the normal suffix rules apply ✨. You can give comma-separated values for reads, writes, and trims, as described in blocksize.

So for example, rate=1m,500k would limit reads to 1MiB/sec and writes to 500KiB/sec. And you can cap just reads or just writes with rate=,500k or rate=500k, — the former will only limit writes (to 500KiB/sec) and the latter will only limit reads. No cap on the other side, she's free 💅.

rate_min=int[,int][,int] — tells fio to do whatever it can to maintain at least this bandwidth. And if it fails to meet this requirement, the job exits 💀. Comma-separated values may be specified for reads, writes, and trims as described in blocksize.
rate_iops=int[,int][,int] — caps the bandwidth to this number of IOPS. Basically the same as rate, just specified independently of bandwidth 🎀. If the job is given a block size range instead of a fixed value, the smallest block size is used as the metric. Comma-separated values may be specified for reads, writes, and trims as described in blocksize.
rate_iops_min=int[,int][,int] — if fio doesn't meet this rate of I/O, it will cause the job to exit 😭. Comma-separated values may be specified for reads, writes, and trims as described in blocksize.
rate_process=str — controls how fio manages rated I/O submissions 🦋. The default is linear, which submits I/O in a linear fashion with fixed delays between I/Os that get adjusted based on I/O completion rates. If you set it to poisson, fio will submit I/O based on a more real-world random request flow, known as the Poisson process (https://en.wikipedia.org/wiki/Poisson_point_process). The lambda will be 10^6 / IOPS for the given workload.
rate_ignore_thinktime=bool — by default, fio will attempt to catch up to the specified rate setting if any kind of thinktime setting was used 🙈. If this option is set, then fio will ignore the thinktime and continue doing IO at the specified rate, instead of entering a catch-up mode after thinktime is done.
rate_cycle=int — average bandwidth for rate_min and rate_iops_min over this number of milliseconds. Defaults to 1000.

💜 I/O latency 💜

latency_target=time — if set, fio will attempt to find the max performance point that the given workload will run at while keeping latency below this target 🎯. When the unit is omitted, the value is interpreted in microseconds. See latency_window and latency_percentile.
latency_window=time — used with latency_target to specify the sample window that the job is run at varying queue depths to test the performance. When the unit is omitted, the value is interpreted in microseconds.
latency_percentile=float — the percentage of I/Os that must fall within the criteria specified by latency_target and latency_window 💖. If not set, this defaults to 100.0, meaning that all I/Os must be equal or below the value set by latency_target. No exceptions, periodt.
latency_run=bool — used with latency_target. If false (that's the default), fio will find the highest queue depth that meets latency_target and exit 👋. If true, fio will continue running and try to meet latency_target by adjusting queue depth.
max_latency=time[,time][,time] — if set, fio will exit the job with an ETIMEDOUT error if it exceeds this maximum latency 💀. When the unit is omitted, the value is interpreted in microseconds. Comma-separated values may be specified for reads, writes, and trims as described in blocksize.

🦋 I/O replay 🦋

write_iolog=str — writes the issued I/O patterns to the specified file ✍️. See read_iolog. Specify a separate file for each job, otherwise the iologs will be interspersed and the file may be corrupt 😭. This file will be opened in append mode.
read_iolog=str — opens an iolog with the specified filename and replays the I/O patterns it contains 🔁. This can be used to store a workload and replay it sometime later, which is lowkey the dream. The iolog given may also be a blktrace binary file, which lets fio replay a workload captured by blktrace. See blktrace(8) for how to capture such logging data. For blktrace replay, the file needs to be turned into a blkparse binary data file first:

blkparse <device> -o /dev/null -d file_for_fio.bin

You can specify a number of files by separating the names with a ':' character. See the filename option for info on how to escape ':' characters within the file names 🫶. These files will be sequentially assigned to job clones created by numjobs. Also, '-' is a reserved name, meaning read from stdin — notably, if filename is set to '-' (which means stdin as well), then this flag can't be set to '-' 🙅‍♀️.

read_iolog_chunked=bool — determines how the iolog is read. If false (the default), the entire read_iolog will be read at once. If you select true, input from the iolog will be read gradually 🌊. Useful when the iolog is very large, or it is generated.
merge_blktrace_file=str — when specified, rather than replaying the logs passed to read_iolog, the logs go through a merge phase which aggregates them into a single blktrace ✨. The resulting file is then passed on as the read_iolog parameter. The intention here is to make the order of events consistent. This limits the influence of the scheduler compared to replaying multiple blktraces via concurrent jobs.
merge_blktrace_scalars=float_list — this is a percentage-based option that is index paired with the list of files passed to read_iolog. When merging is performed, it scales the time of each event by the corresponding amount 🌸. For example, --merge_blktrace_scalars="50:100" runs the first trace in halftime and the second trace in realtime. This knob is separately tunable from replay_time_scale, which scales the trace during runtime and will not change the output of the merge — unlike this option 👀.
merge_blktrace_iters=float_list — this is a whole number option that is index paired with the list of files passed to read_iolog. When merging is performed, it runs each trace for the specified number of iterations 🔂. For example, --merge_blktrace_iters="2:1" runs the first trace for two iterations and the second trace for one iteration.
replay_no_stall=bool — when replaying I/O with read_iolog, the default behavior is to attempt to respect the timestamps within the log and replay them with the appropriate delay between IOPS ⏳. By setting this variable, fio will not respect the timestamps and will attempt to replay them as fast as possible while still respecting ordering. The result is the same I/O pattern to a given device, but different timings 🏃‍♀️💨.
replay_time_scale=int — when replaying I/O with read_iolog, fio will honor the original timing in the trace. With this option, it's possible to scale the time 🎚️. It's a percentage option: if set to 50 it means run at 50% the original IO rate in the trace; if set to 200, run at twice the original IO rate. Defaults to 100.
replay_redirect=str — okay so when you're replaying I/O patterns with read_iolog, the default vibe is that fio replays the IOPS back onto the exact major/minor device each IOP was originally recorded from. Which is sometimes just... not it 💀 because on a different machine those major/minor numbers can totally map to a different device. And even swapping hardware on the same system can give you a different major/minor mapping. That's where replay_redirect comes in: it makes all I/Os replay onto the single device you name, no matter which device they were recorded from. Like, replay_redirect=/dev/sdc would cause all I/O in the blktrace or iolog to be replayed onto /dev/sdc. Which means if your trace contains multiple devices, they're all getting replayed onto one device. If you actually want multiple devices replayed concurrently to multiple redirected devices, you've gotta blkparse your trace into separate traces and replay them with independent fio invocations. And yeah — unfortunately this also breaks the strict time ordering between multiple device accesses. F in chat 😭
replay_align=int — forces the byte offsets in a trace to be aligned to this value. Heads up: the value must be a power of 2. ⚡
replay_scale=int — scales byte offsets down by this factor when replaying traces. You should most likely be using replay_align as well, bestie 🫶

🧵 Threads, processes and job synchronization 🧵

replay_skip=str — sometimes it's genuinely useful to skip certain IO types in a replay trace. Could be nixing the writes in the trace, for instance. Or not replaying the trims/discards if you're redirecting to a device that just doesn't support them 💅 This option takes a comma separated list of read, write, trim, sync.
thread — fio defaults to creating jobs by using fork, but if you give this option, fio will create jobs by using POSIX Threads' function pthread_create(3) to create threads instead. Slay 🧵
wait_for=str — if set, the current job won't even start until all workers of the specified waitee job are done. Now, wait_for operates on the job name basis, so there are a few limitations, and you need to know them:
- First, the waitee must be defined prior to the waiter job — meaning no forward references. No looking into the future, sorry ✨
- Second, if a job is being referenced as a waitee, it must have a unique name (so no duplicate waitees). 👑
comm=str — sets the job process comm to the string you specify. See man prctrl(2). Note: this option is currently supported only on Linux. 🐧
nice=int — runs the job with the given nice value. See man nice(2). On Windows, values less than -15 set the process class to "High"; -1 through -15 set "Above Normal"; 1 through 15 set "Below Normal"; and above 15 gets the "Idle" priority class. Gotta love a platform-specific plot twist 🙈
prio=int — sets the I/O priority value of this job. Linux limits us to a positive value between 0 and 7, with 0 being the highest. See man ionice(1). For other operating systems, refer to an appropriate manpage since the meaning of priority may differ. And for per-command priority setting, see the I/O engine specific cmdprio_percentage and cmdprio options. 🎀
prioclass=int — sets the I/O priority class. See man ionice(1). For per-command priority setting, see the I/O engine specific cmdprio_percentage and cmdprio_class options.
priohint=int — sets the I/O priority hint. This one's only applicable to platforms that support I/O priority classes and to devices with features controlled through priority hints — e.g. block devices supporting command duration limits, or CDL. So like, CDL is a way to indicate the desired maximum latency of I/Os so that the device can optimize its internal command scheduling according to the latency limits indicated by the user. Iconic of the device honestly. For per-I/O priority hint setting, see the I/O engine specific cmdprio_hint option. ✨
cpus_allowed=str — controls the same options as cpumask, but it accepts a textual specification of the permitted CPUs instead, and CPUs are indexed from 0. So if you want CPUs 0 and 5, you'd specify cpus_allowed=0,5. It also lets you do a range of CPUs — say you wanted a binding to CPUs 0, 5, and 8 to 15, you'd set cpus_allowed=0,5,8-15. Chef's kiss range syntax 🤌

On Windows, when cpus_allowed is unset, only CPUs from fio's current processor group will be used, and affinity settings are inherited from the system. An fio build configured to target Windows 7 makes options that set CPUs processor-group aware, and values will set both the processor group and a CPU from within that group. For example, on a system where processor group 0 has 40 CPUs and processor group 1 has 32 CPUs, cpus_allowed values between 0 and 39 will bind CPUs from processor group 0, and cpus_allowed values between 40 and 71 will bind CPUs from processor group 1. When using cpus_allowed_policy=shared, all CPUs specified by a single cpus_allowed option must be from the same processor group. For Windows fio builds not built for Windows 7, CPUs will only be selected from (and be relative to) whatever processor group fio happens to be running in, and CPUs from other processor groups cannot be used. Windows really said "you get this group and that's it" 💀

cpus_allowed_policy=str — sets the policy for how fio distributes the CPUs specified by cpus_allowed or cpumask. Two policies are supported:
- shared — all jobs will share the CPU set specified.
- split — each job will get a unique CPU from the CPU set.

shared is the default behavior if the option isn't specified. If split is specified, then fio will assign one cpu per job. And if there aren't enough CPUs given for the jobs listed, then fio will roundrobin the CPUs in the set. Sharing is caring, but splitting is valid too 🫶

cpumask=int — sets the CPU affinity of this job. The parameter you give is a bit mask of allowed CPUs the job may run on. So if you want the allowed CPUs to be 1 and 5, you'd pass the decimal value of (1 << 1 | 1 << 5), or 34. See man sched_setaffinity(2). Ngl this may not work on all supported operating systems or kernel versions. And it doesn't work well for a higher CPU count than what you can store in an integer mask, so it can only control cpus 1-32. For boxes with larger CPU counts, use cpus_allowed instead. The bit math is giving CS 101 final exam 💀
numa_cpu_nodes=str — sets this job running on specified NUMA nodes' CPUs. The arguments allow a comma delimited list of cpu numbers, A-B ranges, or all. Note: to enable NUMA options support, fio must be built on a system with libnuma-dev(el) installed. 🧚
numa_mem_policy=str — sets this job's memory policy and corresponding NUMA nodes. The format of the arguments is:

<mode>[:<nodelist>]

mode is one of the following memory policies: default, prefer, bind, interleave or local. For default and local memory policies, no node needs to be specified. For prefer, only one node is allowed. For bind and interleave, the nodelist may be as follows: a comma delimited list of numbers, A-B ranges, or all. Pick your fighter ⚔️

cgroup=str — adds the job to this control group. If it doesn't exist, it will be created. The system must have a mounted cgroup blkio mount point for this to work. If your system doesn't have it mounted, you can do so with:

# mount -t cgroup -o blkio none /cgroup

cgroup_weight=int — sets the weight of the cgroup to this value. See the documentation that comes with the kernel — allowed values are in the range of 100..1000. ⚖️
cgroup_nodelete=bool — normally fio will delete the cgroups it has created after the job completion. To override this behavior and leave cgroups around after the job completion, set cgroup_nodelete=1. This can be useful if one wants to inspect various cgroup files after job completion. Default: false. Leaving a little cgroup souvenir behind 🎀
flow_id=int — the ID of the flow. If not specified, it defaults to being a global flow. See flow.
flow=int — weight in token-based flow control. If this value is used, then fio regulates the activity between two or more jobs sharing the same flow_id. Fio attempts to keep each job's activity proportional to other jobs' activities in the same flow_id group, with respect to the requested weight per job. That is, if one job has flow=3, another job has flow=2 and another with flow=1, then there will be a roughly 3:2:1 ratio in how much one runs vs the others. Math but make it fair ⚖️✨
flow_sleep=int — the period of time, in microseconds, to wait after the flow counter has exceeded its proportion before retrying operations. Little cooldown moment 😌
stonewall, wait_for_previous — waits for the preceding jobs in the job file to exit before starting this one. Can be used to insert serialization points in the job file. And a stone wall also implies starting a new reporting group, see group_reporting. Optionally you can use stonewall=0 to disable or stonewall=1 to enable it. Hard boundary bestie 🧱
exitall — by default, fio will continue running all other jobs when one job finishes. Sometimes that's not the desired action. Setting exitall will instead make fio terminate all jobs in the same group as soon as one job of that group finishes. Everybody out 🚪💨
exit_what=str — okay so by default, fio will continue running all other jobs when one job finishes, and sometimes that's not what you want. Setting exitall will instead make fio terminate all jobs in the same group. The option exit_what allows you to control which jobs get terminated when exitall is enabled. The default value is group. The allowed values are:
- all — terminates all jobs.
- group — is the default and does not change the behaviour of exitall.
- stonewall — terminates all currently running jobs across all groups and continues execution with the next stonewalled group.
exec_prerun=str — before running this job, issue the command specified through system(3). Output is redirected in a file called jobname.prerun.txt. Pre-game ritual ✨
exec_postrun=str — after the job completes, issue the command specified though system(3). Output is redirected in a file called jobname.postrun.txt. Post-game debrief 💅

okay so before the verification era, two little identity knobs, bestie:

uid=int — instead of running as the invoking user, set the user ID to this value before the thread/process does any work.
gid=int — set group ID, see uid.

🦋 Verification 🦋

This is where fio checks its own homework, and honestly? iconic behavior. 💅

verify_only — do not perform the specified workload, only verify that data still matches a previous invocation of this workload. This option lets you check data multiple times at a later date without overwriting it. It makes sense only for workloads that write data, and it does not support workloads with the time_based option set. Options verify_write_sequence and verify_header_seed will be disabled in this mode, unless they are explicitly enabled. The writes reported in the output when this option is specified are phantom writes, since no writes are actually issued.
do_verify=bool — run the verify phase after a write phase. Only valid if verify is set. Default: true.
verify=str — if writing to a file, fio can verify the file contents after each iteration of the job. Each verification method also implies verification of a special header, which is written to the beginning of each block. This header also includes meta information, like offset of the block, block number, timestamp when the block was written, initial seed value used to generate the buffer contents, etc. verify can be combined with the verify_pattern option. The allowed values are:
- md5 — use an md5 sum of the data area and store it in the header of each block.
- crc64 — use an experimental crc64 sum of the data area and store it in the header of each block.
- crc32c — use a crc32c sum of the data area and store it in the header of each block. This will automatically use hardware acceleration (e.g. SSE4.2 on an x86 or CRC crypto extensions on ARM64) but will fall back to software crc32c if none is found. Generally the fastest checksum fio supports when hardware accelerated.
- crc32c-intel — synonym for crc32c.
- crc32 — use a crc32 sum of the data area and store it in the header of each block.
- crc16 — use a crc16 sum of the data area and store it in the header of each block.
- crc7 — use a crc7 sum of the data area and store it in the header of each block.
- xxhash — use xxhash as the checksum function. Generally the fastest software checksum that fio supports.
- sha512 — use sha512 as the checksum function.
- sha256 — use sha256 as the checksum function.
- sha1 — use optimized sha1 as the checksum function.
- sha3-224 — use optimized sha3-224 as the checksum function.
- sha3-256 — use optimized sha3-256 as the checksum function.
- sha3-384 — use optimized sha3-384 as the checksum function.
- sha3-512 — use optimized sha3-512 as the checksum function.
- meta — this option is deprecated, since now meta information is included in the generic verification header and meta verification happens by default. For detailed information see the description of the verify setting. This option is kept because of compatibility's sake with old configurations. Do not use it. 💀
- pattern — verify a strict pattern. Normally fio includes a header with some basic information and checksumming, but if this option is set, only the specific pattern set with verify_pattern is verified.
- pattern_hdr — verify a pattern in conjunction with a header.
- null — only pretend to verify. Useful for testing internals with ioengine=null, not for much else.

This option can be used for repeated burn-in tests of a system to make sure that the written data is also correctly read back. 🔥

If the data direction given is a read or random read, fio will assume that it should verify a previously written file. In this scenario fio will not verify the block number written in the header. The header seed won't be verified, unless it's explicitly requested by setting the verify_header_seed option. Note in this scenario the header seed check will only work if the read invocation exactly matches the original write invocation.

If the data direction includes any form of write, the verify will be of the newly written data.

⚠️ real talk: to avoid false verification errors, do not use the norandommap option when verifying data with async I/O engines and I/O depths > 1. Or use the norandommap and the lfsr random generator together to avoid writing to the same offset with multiple outstanding I/Os. no cap.

verify_offset=int — swap the verification header with data somewhere else in the block before writing. It is swapped back before verifying. This should be within the range of verify_interval.
verify_interval=int — write the verification header at a finer granularity than the blocksize. It will be written for chunks the size of verify_interval. blocksize should divide this evenly.
verify_pattern=str — if set, fio will fill the I/O buffers with this pattern. Fio defaults to filling with totally random bytes, but sometimes it's interesting to fill with a known pattern for I/O verification purposes. Depending on the width of the pattern, fio will fill 1/2/3/4 bytes of the buffer at the time (it can be either a decimal or a hex number). The verify_pattern if larger than a 32-bit quantity has to be a hex number that starts with either "0x" or "0X". Use with verify. Also, verify_pattern supports %o format, which means that for each block the offset will be written and then verified back, e.g.:

verify_pattern=%o

Or use combination of everything:

verify_pattern=0xff%o"abcd"-12

verify_pattern_interval=bool — recreate an instance of the verify_pattern every verify_pattern_interval bytes. This is only useful when verify_pattern contains the %o format specifier and can be used to speed up the process of writing each block on a device with its offset. Default: 0 (disabled).
verify_fatal=bool — normally fio will keep checking the entire contents before quitting on a block verification failure. If this option is set, fio will exit the job on the first observed failure. Default: false.
verify_dump=bool — if set, dump the contents of both the original data block and the data block we read off disk to files. This allows later analysis to inspect just what kind of data corruption occurred. Off by default.
verify_async=int — fio will normally verify I/O inline from the submitting thread. This option takes an integer describing how many async offload threads to create for I/O verification instead, causing fio to offload the duty of verifying I/O contents to one or more separate threads. If using this offload option, even sync I/O engines can benefit from using an iodepth setting higher than 1, as it allows them to have I/O in flight while verifies are running. Defaults to 0 async threads, i.e. verification is not asynchronous.
verify_async_cpus=str — tell fio to set the given CPU affinity on the async I/O verification threads. See cpus_allowed for the format used.
verify_backlog=int — fio will normally verify the written contents of a job that utilizes verify once that job has completed. In other words, everything is written then everything is read back and verified. You may want to verify continually instead for a variety of reasons. Fio stores the meta data associated with an I/O block in memory, so for large verify workloads, quite a bit of memory would be used up holding this meta data. If this option is enabled, fio will write only N blocks before verifying these blocks.
verify_backlog_batch=int — control how many blocks fio will verify if verify_backlog is set. If not set, will default to the value of verify_backlog (meaning the entire queue is read back and verified). If verify_backlog_batch is less than verify_backlog then not all blocks will be verified, if verify_backlog_batch is larger than verify_backlog, some blocks will be verified more than once.
verify_state_save=bool — when a job exits during the write phase of a verify workload, save its current state. This allows fio to replay up until that point, if the verify state is loaded for the verify read phase. The format of the filename is, roughly:

<type>-<jobname>-<jobindex>-verify.state.

<type> is "local" for a local run, "sock" for a client/server socket connection, and "ip" (192.168.0.1, for instance) for a networked client/server connection. Defaults to true.

verify_state_load=bool — if a verify termination trigger was used, fio stores the current write state of each thread. This can be used at verification time so that fio knows how far it should verify. Without this information, fio will run a full verification pass, according to the settings in the job file used. Default false.
experimental_verify=bool — enable experimental verification. Standard verify records I/O metadata for later use during the verification phase. Experimental verify instead resets the file after the write phase and then replays I/Os for the verification phase. 🧚
verify_write_sequence=bool — this one checks the header write sequence number. Heads up tho: if you've got multiple jobs running, the write sequence number check can legit fail 😭. Turning this off means fio just skips that check entirely. Why would you do that? Because it's actually useful for testing atomic writes — you still get checksum verification attempted. And when atomic is enabled, checksum verification is expected to succeed (meanwhile the write sequence checking can still fail). So yeah, it's giving "let me test one thing at a time" 💅.
verify_header_seed=bool — verifies the header seed value that was used to generate the buffer contents. In certain scenarios like read/verify-only workloads, when norandommap is enabled, or with offset modifiers (go look at the readwrite and rw_sequencer options), the header seed verification can fail. Disabling this means header seed checking gets skipped. Defaults to true. 🫶
trim_percentage=int — how many verify blocks to discard/trim.
trim_verify_zero=bool — verifies that trim/discarded blocks come back as zeros.
trim_backlog=int — verifies that trim/discarded blocks come back as zeros. (yes bestie, same wording as the one above — that's what the page says, I don't make the rules 🙈)
trim_backlog_batch=int — trim this number of I/O blocks.

✨ Steady state ✨

steadystate=str:float, ss=str:float — this is how you define the criterion and the limit for assessing steady state performance. First parameter = the criterion, second parameter = the threshold. When the criterion falls below the threshold for the specified duration, the job stops. Example time: iops_slope:0.1% tells fio to terminate the job when the least squares regression slope falls below 0.1% of the mean IOPS. If group_reporting is enabled, this applies to all jobs in the group. Below is the list of available steady state assessment criteria — and all assessments are carried out using only data from the rolling collection window. Threshold limits can be expressed as a fixed value or as a percentage of the mean in the collection window. 🎀

When you're using this feature, most jobs should include the time_based and runtime options, or the loops option, so fio doesn't just stop running after it's covered the full size of the specified file(s) or device(s). No cap, that one matters.

iops — collect IOPS data. Stop the job if all individual IOPS measurements are within the specified limit of the mean IOPS (e.g., iops:2 means all individual IOPS values must be within 2 of the mean, whereas iops:0.2% means all individual IOPS values must be within 0.2% of the mean IOPS to terminate the job).
iops_slope — collect IOPS data and calculate the least squares regression slope. Stop the job if the slope falls below the specified limit.
bw — collect bandwidth data. Stop the job if all individual bandwidth measurements are within the specified limit of the mean bandwidth.
bw_slope — collect bandwidth data and calculate the least squares regression slope. Stop the job if the slope falls below the specified limit.

steadystate_duration=time, ss_dur=time — a rolling window of this duration gets used to judge whether steady state has been reached. Data is collected every ss_interval. The default is 0, which disables steady state detection. When the unit is omitted, the value is interpreted in seconds. ⭐
steadystate_ramp_time=time, ss_ramp=time — lets the job run for the specified duration before it starts collecting data for checking the steady state job termination criterion. The default is 0. When the unit is omitted, the value is interpreted in seconds. 🌸
steadystate_check_interval=time, ss_interval=time — the values during the rolling window get collected with a period of this value. If ss_interval is 30s and ss_dur is 300s, you get 10 measurements taken. Default is 1s, but that might not converge — especially for slower devices — so set this accordingly. When the unit is omitted, the value is interpreted in seconds. 🦋

✨ Measurements and reporting ✨

per_job_logs=bool — if set to true, fio generates bw/clat/iops logs with per-job unique filenames. If set to false, jobs with identical names will share a log filename. Note that when this is false, log files get opened in append mode, and if log files already exist, the previous contents will not be overwritten. Default: true. 👑
group_reporting — sometimes it's way more interesting to display statistics for groups of jobs as a whole instead of for each individual job. This is especially true if numjobs is used — staring at individual thread/process output quickly becomes unwieldy, like 😭. To see the final report per-group instead of per-job, use group_reporting. Jobs in a file will be part of the same reporting group, unless they're separated by a stonewall, or by using new_group.

NOTE: When group_reporting is used along with json output, there are certain per-job properties which can be different between jobs but don't have a natural group-level equivalent. Examples include kb_base, unit_base, sig_figs, thread_number, pid, and job_start. For these properties, the values for the first job get recorded for the group. iykyk.

Also, options like percentile_list and unified_rw_reporting should be consistent among the jobs in a reporting group. Having options like these vary across the jobs in a reporting group is an unsupported configuration. Just... don't 💀.

new_group — starts a new reporting group. See: group_reporting. If it's not given, all jobs in a file will be part of the same reporting group, unless separated by a stonewall.
stats=bool — by default, fio collects and shows final output results for all jobs that run. If this option is set to 0, then fio will ignore it in the final stat output.
write_bw_log=str — if given, writes a bandwidth log for this job. Can be used to store data of the bandwidth of the jobs in their lifetime. 📈

If no str argument is given, the default filename of jobname_type.x.log is used. Even when the argument is given, fio will still append the type of log. So if you specify:

write_bw_log=foo

the actual log name will be foo_bw.x.log, where x is the index of the job (1..N, where N is the number of jobs). If per_job_logs is false, then the filename will not include the .x job index.

The included fio_generate_plots script uses gnuplot to turn these text files into nice graphs. See the LOG FILE FORMATS section for how data is structured within the file. ✨

write_lat_log=str — same as write_bw_log, except this option creates I/O submission (e.g., name_slat.x.log), completion (e.g., name_clat.x.log), and total (e.g., name_lat.x.log) latency files instead. See write_bw_log for details about the filename format, and the LOG FILE FORMATS section for how data is structured within the files. 🎀
write_hist_log=str — same as write_bw_log but writes an I/O completion latency histogram file (e.g., name_hist.x.log) instead. Note that this file will be empty unless log_hist_msec has also been set. See write_bw_log for details about the filename format, and the LOG FILE FORMATS section for how data is structured within the file.
write_iops_log=str — same as write_bw_log, but writes an IOPS file (e.g. name_iops.x.log) instead. Because fio defaults to individual I/O logging, the value entry in the IOPS log will be 1 unless windowed logging (see log_avg_msec) has been enabled. See write_bw_log for details about the filename format, and LOG FILE FORMATS for how data is structured within the file. 💖
log_entries=int — by default, fio will log an entry in the iops, latency, or bw log for every I/O that completes. The initial number of I/O log entries is 1024. When the log entries are all used, new log entries are dynamically allocated. This dynamic log entry allocation may negatively impact time-related statistics such as I/O tail latencies (e.g. 99.9th percentile completion latency). This option lets you specify a larger initial number of log entries to avoid run-time allocation of new log entries, resulting in more precise time-related I/O statistics. Also see log_avg_msec as well. Defaults to 1024. 🔥
log_avg_msec=int — by default, fio logs an entry in the iops, latency, or bw log for every I/O that completes. When writing to the disk log, that can quickly grow to a very large size 😳. Setting this option directs fio to instead record an average over the specified duration for each log entry, reducing the resolution of the log. When the job completes, fio will flush any accumulated latency log data, so the final log interval may not match the value specified by this option — and there may even be duplicate timestamps. See log_window_value as well. Defaults to 0, logging entries for each I/O. Also see the LOG FILE FORMATS section. 🍓
log_hist_msec=int — same as log_avg_msec, but logs entries for completion latency histograms. Computing latency percentiles from averages of intervals using log_avg_msec is inaccurate. Setting this option makes fio log histogram entries over the specified period of time, reducing log sizes for high IOPS devices while retaining percentile accuracy. See log_hist_coarseness and write_hist_log as well. Defaults to 0, meaning histogram logging is disabled. 💜
log_hist_coarseness=int — an integer from 0 to 6 that sets how coarse the resolution is for the histogram logs you turned on with log_hist_msec. every time coarseness goes up by one, fio spits out half as many bins 🎚️. defaults to 0, and at 0 your histogram logs have 1216 latency bins. peep the LOG FILE FORMATS section for the deets.
log_window_value=str, log_max_value=str — so if log_avg_msec is set, fio's default behavior is logging the average over that window. this option decides whether fio logs the average, the max, or literally both over the window 🔥. it only affects latency logging, since the average and max values for iops or bw logs are the same anyway. the values it accepts are:
- avg — log the average value over the window. this is the default.
- max — log the maximum value in the window.
- both — log both the average and the maximum over the window.
- 0 — a backward-compatible alias for avg.
- 1 — a backward-compatible alias for max.
log_offset=bool — set this and the iolog options will include the byte offset for the I/O entry alongside the other data values. defaults to 0, which means nope, no offsets in your logs. also go look at LOG FILE FORMATS 👀.
log_prio=bool — if this is set, the Command priority field in LOG FILE FORMATS shows the priority value and the IO priority class of the command. otherwise, that field just shows whether the command has the highest RT priority class or not. again, LOG FILE FORMATS section, go read it 💅.
log_issue_time=bool — set this and the iolog options include the command issue time for the I/O entry along with the other data values. defaults to 0, meaning command issue times are not in the logs. see LOG FILE FORMATS. > heads up bestie: this option has to be set together with write_lat_log and log_offset, mkay?
log_compression=int — set this and fio compresses the I/O logs as it goes so the memory footprint stays lower 🫶. when a log hits the size you specified, that chunk gets removed and compressed in the background. I/O logs compress really well, so this gives you a genuinely nice memory savings on longer runs. the downside? that compression eats some background CPU cycles, so it may impact the run. but tbh that's also true if the logging ends up eating most of your system memory — so pick your poison ☠️. the I/O logs get saved normally at the end of the run: fio decompresses the chunks and stores them in the log file you specified. this whole feature depends on whether zlib is available.
log_compression_cpus=str — defines the set of CPUs allowed to handle online log compression for the I/O jobs. this can give you better isolation between your performance-sensitive jobs and the background compression work ✨. for the format, check out cpus_allowed.
log_store_compressed=bool — if set, fio stores the log files in a compressed format. you can decompress them with fio using the --inflate-log command line parameter. the files get stored with a .fz suffix 💾.
log_unix_epoch=bool — backward-compatible alias for log_alternate_epoch.
log_alternate_epoch=bool — if set, fio logs timestamps based on the epoch used by the clock specified in the log_alternate_epoch_clock_id option, into the log files produced by enabling write_type_log for each log type, instead of the default zero-based timestamps 🕰️.
log_alternate_epoch_clock_id=int — specifies the clock_id that clock_gettime uses to obtain the alternate epoch, if log_alternate_epoch is true. otherwise it does nothing at all. default value is 0, aka CLOCK_REALTIME.
block_error_percentiles=bool — if set, record errors in trim block-sized units from writes and trims, then output a histogram of how many trims it took to get to the errors and what kind of error was encountered 📊.
bwavgtime=int — average the calculated bandwidth over the given time. the value is in milliseconds ⏱️. if the job also does bandwidth logging through write_bw_log, then the minimum of this option and log_avg_msec gets used. default: 500ms.
iopsavgtime=int — average the calculated IOPS over the given time. value in milliseconds. if the job also does IOPS logging through write_iops_log, then the minimum of this option and log_avg_msec is used. default: 500ms.
disk_util=bool — generate disk utilization statistics, if the platform supports it. default: true.
disable_lat=bool — disables measurements of total latency numbers. only actually useful for cutting back the number of calls to gettimeofday(2), because those calls genuinely do impact performance at really high IOPS rates. > note: to really get rid of a large amount of these calls, you gotta use this option together with disable_slat and disable_bw_measurement too.
disable_clat=bool — disables measurements of completion latency numbers. see disable_lat above 👆.
disable_slat=bool — disables measurements of submission latency numbers. see disable_lat.
disable_bw_measurement=bool, disable_bw=bool — disables measurements of throughput/bandwidth numbers. see disable_lat.
slat_percentiles=bool — report submission latency percentiles. btw submission latency is not recorded for synchronous ioengines, so yeah, keep that in mind 🫠.
clat_percentiles=bool — report completion latency percentiles.
lat_percentiles=bool — report total latency percentiles. total latency is the sum of submission latency plus completion latency 🧮.
percentile_list=float_list — overwrites the default list of percentiles for latencies and for the block error histogram. every number is a floating point number in the range (0,100], and the max length of the list is 20. separate the numbers with ':'. for example, --percentile_list=99.5:99.9' makes fio report the latency durations below which 99.5% and 99.9% of the observed latencies fell, respectively. iconic, honestly.
significant_figures=int — if you're using --output-format of normal, set the significant figures to this value. higher values give you more precise IOPS and throughput units, lower values round. needs a minimum value of 1 and a maximum value of 10. defaults to 4. ✨

🚨 Error handling 🚨

exitall_on_error — when one job finishes in error, terminate the rest. the default is to wait for each job to finish.
continue_on_error=str — normally fio exits the job on the first failure it observes. if you set this option, fio keeps the job going when there's a 'non-fatal error' (EIO or EILSEQ) until the runtime is exceeded or the I/O size you specified is completed 🏃‍♀️. if you use this option, there are two more stats appended: the total error count and the first error. the error field given in the stats is the first error that got hit during the run.

Note: a write error from the device may legitimately go unnoticed by fio when you're using buffered IO, because the write() (or similar) system call merely dirties the kernel pages — unless sync or direct is used. Device IO errors happen when the dirty data is actually written out to disk. if fully sync writes aren't what you want, fsync or fdatasync can be used as well. this is specific to writes, since reads are always synchronous. 😭

the allowed values are:

none — exit on any I/O or verify errors.
read — continue on read errors, exit on all others.
write — continue on write errors, exit on all others.
io — continue on any I/O error, exit on all others.
verify — continue on verify errors, exit on all others.
all — continue on all errors.
0 — backward-compatible alias for 'none'.
1 — backward-compatible alias for 'all'.

ignore_error=str — sometimes you just wanna ignore some errors during a test. in that case you can specify an error list for each error type, instead of only being able to ignore the default 'non-fatal error' via continue_on_error. the format is ignore_error=READ_ERR_LIST,WRITE_ERR_LIST,VERIFY_ERR_LIST, and the errors for a given error type are separated with ':'. an error may be a symbol ('ENOSPC', 'ENOMEM') or an integer. example:

ignore_error=EAGAIN,ENOSPC:122

this option will ignore EAGAIN from READ, and ENOSPC and 122(EDQUOT) from WRITE. it works by overriding continue_on_error with the list of errors for each error type, if any 🎀.

error_dump=bool — if set, dump every error even if it's non-fatal. true by default. if disabled, only fatal errors get dumped.

🏃‍♀️ Running predefined workloads 🏃‍♀️

fio includes predefined profiles that mimic the I/O workloads generated by other tools 👑.

profile=str — the predefined workload to run. current profiles are:
- tiobench — Threaded I/O bench (tiotest/tiobench) like workload.
- act — Aerospike Certification Tool (ACT) like workload.

to view a profile's additional options, use --cmdhelp after specifying the profile. for example:

$ fio --profile=act --cmdhelp

🌈 Act profile options 🌈

device-names=str — devices to use.
load=int — ACT load multiplier. default: 1.
test-duration=time — how long the entire test takes to run. when the unit is omitted, the value is given in seconds. default: 24h.
threads-per-queue=int — number of read I/O threads per device. default: 8.
read-req-num-512-blocks=int — number of 512B blocks to read at the time. default: 3.
large-block-op-kbytes=int — size of large block ops in KiB (writes). default: 131072.
prep — set to run ACT prep phase. 💅

🌸 Tiobench profile options 🌸

size=str — the size, in MiB.
block=int — block size in bytes. Default: 4096.
numruns=int — how many runs.
dir=str — the test directory.
threads=int — number of threads.

✨ OUTPUT ✨

okay bestie, fio is a yapper. It spits out a lot of output. While it's running, fio shows you the status of the jobs it made. Example time:

          Jobs: 1 (f=1): [_(1),M(1)][24.8%][r=20.5MiB/s,w=23.5MiB/s][r=82,w=94 IOPS][eta 01m:31s]

The characters inside that first set of square brackets are the current status of each thread. First character = first job defined in the job file, and so forth. The possible values, in typical life-cycle order, are:

P — thread setup, but not started.
C — thread created.
I — thread initialized, waiting or generating necessary data.
p — thread running pre-reading file(s).
/ — thread is in ramp period.
R — running, doing sequential reads.
r — running, doing random reads.
W — running, doing sequential writes.
w — running, doing random writes.
M — running, doing mixed sequential reads/writes.
m — running, doing mixed random reads/writes.
D — running, doing sequential trims.
d — running, doing random trims.
F — running, currently waiting for fsync(2).
V — running, doing verification of written data.
f — thread finishing.
E — thread exited, not reaped by main thread yet.
- — thread reaped.
X — thread reaped, exited with an error.
K — thread reaped, exited due to signal.

Fio condenses that thread string so it doesn't hog more space on the command line than it needs to. Like, if you've got 10 readers and 10 writers going, the output looks like this:

          Jobs: 20 (f=20): [R(10),W(10)][4.0%][r=20.5MiB/s,w=23.5MiB/s][r=82,w=94 IOPS][eta 57m:36s]

Note that the status string is displayed in order, so you can literally tell which jobs are doing what. In the example above, that means jobs 1--10 are readers and 11--20 are writers. iykyk 💅

The other values are pretty self-explanatory — number of threads currently running and doing I/O, the number of currently open files (f=), the estimated completion percentage, the rate of I/O since the last check (read speed listed first, then write speed, and optionally trim speed) in terms of bandwidth and IOPS, and the time to completion for the current running group. It's impossible to estimate the runtime of the following groups (if any). Sad but true 😭

When fio is done (or interrupted by Ctrl+C), it will show the data for each thread, group of threads, and disks, in that order. For each overall thread (or group), the output looks like:

          Client1: (groupid=0, jobs=1): err= 0: pid=16109: Sat Jun 24 12:07:54 2017
            write: IOPS=88, BW=623KiB/s (638kB/s)(30.4MiB/50032msec)
              slat (nsec): min=500, max=145500, avg=8318.00, stdev=4781.50
              clat (usec): min=170, max=78367, avg=4019.02, stdev=8293.31
               lat (usec): min=174, max=78375, avg=4027.34, stdev=8291.79
              clat percentiles (usec):
               |  1.00th=[  302],  5.00th=[  326], 10.00th=[  343], 20.00th=[  363],
               | 30.00th=[  392], 40.00th=[  404], 50.00th=[  416], 60.00th=[  445],
               | 70.00th=[  816], 80.00th=[ 6718], 90.00th=[12911], 95.00th=[21627],
               | 99.00th=[43779], 99.50th=[51643], 99.90th=[68682], 99.95th=[72877],
               | 99.99th=[78119]
             bw (  KiB/s): min=  532, max=  686, per=0.10%, avg=622.87, stdev=24.82, samples=  100
             iops        : min=   76, max=   98, avg=88.98, stdev= 3.54, samples=  100
            lat (usec)   : 250=0.04%, 500=64.11%, 750=4.81%, 1000=2.79%
            lat (msec)   : 2=4.16%, 4=1.84%, 10=4.90%, 20=11.33%, 50=5.37%
            lat (msec)   : 100=0.65%
            cpu          : usr=0.27%, sys=0.18%, ctx=12072, majf=0, minf=21
            IO depths    : 1=85.0%, 2=13.1%, 4=1.8%, 8=0.1%, 16=0.0%, 32=0.0%, >=64=0.0%
               submit    : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
               complete  : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
               issued rwt: total=0,4450,0, short=0,0,0, dropped=0,0,0
               latency   : target=0, window=0, percentile=100.00%, depth=8

The job name (or the first job's name when using group_reporting) gets printed, along with the group id, the count of jobs being aggregated, the last error id seen (which is 0 when there are no errors), the pid/tid of that thread, and the time the job/group completed. Below that are the I/O statistics for each data direction performed (writes, in the example above). In the order listed, they mean:

read/write/trim — the string before the colon shows the I/O direction the stats are for. IOPS is the average I/Os performed per second. BW is the average bandwidth rate shown as: value in power of 2 format (value in power of 10 format). The last two values show: (total I/O performed in power of 2 format / runtime of that thread).
slat — submission latency (min being the minimum, max being the maximum, avg being the average, stdev being the standard deviation). This is the time it took to submit the I/O. For sync I/O this row is not displayed, because the slat is really the completion latency (since queue/complete is one operation there). This value can be in nanoseconds, microseconds or milliseconds — fio picks the most appropriate base and prints that (in the example above, nanoseconds was the best scale). Note: in --minimal mode latencies are always expressed in microseconds.
clat — completion latency. Same names as slat, this is the time from submission to completion of the I/O pieces. For sync I/O, clat will usually be equal (or very close) to 0, since the time from submit to complete is basically just CPU time (the I/O has already been done, see the slat explanation).

For file and directory operation engines, clat denotes the time to complete one file or directory operation.

filecreate engine: the time cost to create a new file

filestat engine: the time cost to look up an existing file

filedelete engine: the time cost to delete a file

dircreate engine: the time cost to create a new directory

dirstat engine: the time cost to look up an existing directory

dirdelete engine: the time cost to delete a directory

lat — total latency. Same names as slat and clat, this is the time from when fio created the I/O unit to completion of the I/O operation.
bw — bandwidth statistics based on measurements from discrete intervals. Fio continuously monitors bytes transferred and I/O operations completed. By default fio calculates bandwidth in each half-second interval (see bwavgtime) and reports descriptive statistics for the measurements here. Same names as the xlat stats, but also includes the number of samples taken (samples) and an approximate percentage of total aggregate bandwidth this thread received in its group (per). That last value is only really useful if the threads in this group are on the same disk, since they're then competing for disk access.

For file and directory operation engines, bw is meaningless.

iops — IOPS statistics based on measurements from discrete intervals. For details see the description for bw above. See iopsavgtime to control the duration of the intervals. Same values reported here as for bw, except for percentage.

For file and directory operation engines, iops is the most fundamental index to denote the performance. It means how many files or directories can be operated per second.

filecreate engine: number of files can be created per second

filestat engine: number of files can be looked up per second

filedelete engine: number of files can be deleted per second

dircreate engine: number of directories can be created per second

dirstat engine: number of directories can be looked up per second

dirdelete engine: number of directories can be deleted per second

lat (nsec/usec/msec) — the distribution of I/O completion latencies. This is the time from when I/O leaves fio to when it gets completed. Unlike the separate read/write/trim sections above, the data here and in the remaining sections apply to all I/Os for the reporting group. 250=0.04% means that 0.04% of the I/Os completed in under 250us. 500=64.11% means that 64.11% of the I/Os required 250 to 499us for completion.
cpu — CPU usage. User and system time, along with the number of context switches this thread went through, usage of system and user time, and finally the number of major and minor page faults. The CPU utilization numbers are averages for the jobs in that reporting group, while the context and fault counters are summed.
IO depths — the distribution of I/O depths over the job lifetime. The numbers are divided into powers of 2, and each entry covers depths from that value up to those that are lower than the next entry — e.g. 16= covers depths from 16 to 31. Note that the range covered by a depth distribution entry can be different to the range covered by the equivalent submit/complete distribution entry.

IO submit 🎀

How many pieces of I/O got submitted in a single submit call, bestie. Each entry means "that amount and below, until the previous entry" — so like, 16=100% means we submitted anywhere between 9 and 16 I/Os per submit call. ⭐ Heads up though: the range covered by a submit distribution entry can legitimately be different from the range covered by the equivalent depth distribution entry. Not a bug, just how it is. 🙈

IO complete 💅

Same vibe as the submit number above, except it's tracking completions instead. 🦋

IO issued rwt 🔥

This one's the number of read/write/trim requests issued, plus how many of them were short or dropped. 👑

IO latency ✨

These values are for latency_target and its related options. When those options are engaged, this section tells you the I/O depth required to meet the specified latency target. It's giving "how deep do we gotta go to hit the goal". 😭

After each client has been listed, the group statistics get printed. They look like this:

Run status group 0 (all jobs):
   READ: bw=20.9MiB/s (21.9MB/s), 10.4MiB/s-10.8MiB/s (10.9MB/s-11.3MB/s), io=64.0MiB (67.1MB), run=2973-3069msec
  WRITE: bw=1231KiB/s (1261kB/s), 616KiB/s-621KiB/s (630kB/s-636kB/s), io=64.0MiB (67.1MB), run=52747-53223msec

For each data direction it prints:

bw — the aggregate bandwidth of threads in this group, followed by the minimum and maximum bandwidth of all the threads in this group. Values outside the brackets are power-of-2 format, and the ones within are the equivalent value in power-of-10 format. 🫶
io — the aggregate I/O performed by all threads in this group. Same format as bw, bestie.
run — the smallest and longest runtimes of the threads in this group. ⏱️

And finally, the disk statistics get printed. This is Linux specific, so don't come for me if you're elsewhere. 💀 They look like this:

Disk stats (read/write):
  sda: ios=16398/16511, sectors=32321/65472, merge=30/162, ticks=6853/819634, in_queue=826487, util=100.00%

Each value is printed for both reads and writes — reads first, always. 🌸 The numbers mean:

ios — the number of I/Os performed by all groups.
merge — the number of merges performed by the I/O scheduler.
ticks — the number of ticks we kept the disk busy.
in_queue — total time spent in the disk queue.
util — the disk utilization. 100% means we kept the disk busy constantly, and 50% would be a disk idling half of the time. Iconic little metric, ngl. 👑

It is also possible to get fio to dump the current output while it's running, without terminating the job. To do that, send fio the USR1 signal. 🎀 You can also get regularly timed dumps by using the --status-interval parameter, or by creating a file in /tmp named fio-dump-status. If fio sees this file, it will unlink it and dump the current output status. Main character behavior fr fr. 💅

🌈 TERSE OUTPUT 🌈

For scripted usage, where you typically want to generate tables or graphs of the results, fio can output the results in a semicolon separated format. The format is one long line of values, such as:

2;card0;0;0;7139336;121836;60004;1;10109;27.932460;116.933948;220;126861;3495.446807;1085.368601;226;126864;3523.635629;1089.012448;24063;99944;50.275485%;59818.274627;5540.657370;7155060;122104;60004;1;8338;29.086342;117.839068;388;128077;5032.488518;1234.785715;391;128085;5061.839412;1236.909129;23436;100928;50.287926%;59964.832030;5644.844189;14.595833%;19.394167%;123706;0;7313;0.1%;0.1%;0.1%;0.1%;0.1%;0.1%;100.0%;0.00%;0.00%;0.00%;0.00%;0.00%;0.00%;0.01%;0.02%;0.05%;0.16%;6.04%;40.40%;52.68%;0.64%;0.01%;0.00%;0.01%;0.00%;0.00%;0.00%;0.00%;0.00%
A description of this job goes here.

The job description (if provided) follows on a second line for terse v2. It appears on the same line for other terse versions. Don't ask, it's just the lore. 😭

To enable terse output, use the --minimal or --output-format=terse command line options. The first value is the version of the terse output format. If the output has to be changed for some reason, this number gets incremented by 1 to signify that change. ⭐

Split up, the format is as follows (comments in brackets denote when a field was introduced or whether it's specific to some terse version):

       terse version, fio version [v3], jobname, groupid, error

READ status:

       Total IO (KiB), bandwidth (KiB/sec), IOPS, runtime (msec)
       Submission latency: min, max, mean, stdev (usec)
       Completion latency: min, max, mean, stdev (usec)
       Completion latency percentiles: 20 fields (see below)
       Total latency: min, max, mean, stdev (usec)
       Bw (KiB/s): min, max, aggregate percentage of total, mean, stdev, number of samples [v5]
       IOPS [v5]: min, max, mean, stdev, number of samples

WRITE status:

       Total IO (KiB), bandwidth (KiB/sec), IOPS, runtime (msec)
       Submission latency: min, max, mean, stdev (usec)
       Completion latency: min, max, mean, stdev (usec)
       Completion latency percentiles: 20 fields (see below)
       Total latency: min, max, mean, stdev (usec)
       Bw (KiB/s): min, max, aggregate percentage of total, mean, stdev, number of samples [v5]
       IOPS [v5]: min, max, mean, stdev, number of samples

TRIM status [all but version 3]:

       Fields are similar to READ/WRITE status.

CPU usage:

       user, system, context switches, major faults, minor faults

I/O depths:

       <=1, 2, 4, 8, 16, 32, >=64

I/O latencies microseconds:

       <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000

I/O latencies milliseconds:

       <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000, 2000, >=2000

Disk utilization [v3]:

       disk name, read ios, write ios, read merges, write merges, read ticks, write ticks, time spent in queue, disk utilization percentage

Additional Info (dependent on continue_on_error, default off):

       total # errors, first error code

Additional Info (dependent on description being set):

       Text description

Completion latency percentiles can be a grouping of up to 20 sets, so for the terse output fio writes all of them. Each field will look like this:

1.00%=6112

which is the Xth percentile, and the usec latency associated with it. 🎀

For Disk utilization, all disks used by fio are shown. So for each disk there will be a disk utilization section. Periodt. 💅

Below is a single line containing short names for each of the fields in the minimal output v3, separated by semicolons:

terse_version_3;fio_version;jobname;groupid;error;read_kb;read_bandwidth_kb;read_iops;read_runtime_ms;read_slat_min_us;read_slat_max_us;read_slat_mean_us;read_slat_dev_us;read_clat_min_us;read_clat_max_us;read_clat_mean_us;read_clat_dev_us;read_clat_pct01;read_clat_pct02;read_clat_pct03;read_clat_pct04;read_clat_pct05;read_clat_pct06;read_clat_pct07;read_clat_pct08;read_clat_pct09;read_clat_pct10;read_clat_pct11;read_clat_pct12;read_clat_pct13;read_clat_pct14;read_clat_pct15;read_clat_pct16;read_clat_pct17;read_clat_pct18;read_clat_pct19;read_clat_pct20;read_tlat_min_us;read_lat_max_us;read_lat_mean_us;read_lat_dev_us;read_bw_min_kb;read_bw_max_kb;read_bw_agg_pct;read_bw_mean_kb;read_bw_dev_kb;write_kb;write_bandwidth_kb;write_iops;write_runtime_ms;write_slat_min_us;write_slat_max_us;write_slat_mean_us;write_slat_dev_us;write_clat_min_us;write_clat_max_us;write_clat_mean_us;write_clat_dev_us;write_clat_pct01;write_clat_pct02;write_clat_pct03;write_clat_pct04;write_clat_pct05;write_clat_pct06;write_clat_pct07;write_clat_pct08;write_clat_pct09;write_clat_pct10;write_clat_pct11;write_clat_pct12;write_clat_pct13;write_clat_pct14;write_clat_pct15;write_clat_pct16;write_clat_pct17;write_clat_pct18;write_clat_pct19;write_clat_pct20;write_tlat_min_us;write_lat_max_us;write_lat_mean_us;write_lat_dev_us;write_bw_min_kb;write_bw_max_kb;write_bw_agg_pct;write_bw_mean_kb;write_bw_dev_kb;cpu_user;cpu_sys;cpu_csw;cpu_mjf;cpu_minf;iodepth_1;iodepth_2;iodepth_4;iodepth_8;iodepth_16;iodepth_32;iodepth_64;lat_2us;lat_4us;lat_10us;lat_20us;lat_50us;lat_100us;lat_250us;lat_500us;lat_750us;lat_1000us;lat_2ms;lat_4ms;lat_10ms;lat_20ms;lat_50ms;lat_100ms;lat_250ms;lat_500ms;lat_750ms;lat_1000ms;lat_2000ms;lat_over_2000ms;disk_name;disk_read_iops;disk_write_iops;disk_read_merges;disk_write_merges;disk_read_ticks;write_ticks;disk_queue_time;disk_util

In client/server mode terse output differs from what appears when jobs are run locally. 🙈 Disk utilization data is omitted from the standard terse output, and for v3 and later it appears on its own separate line at the end of each terse reporting cycle.

🦋 JSON OUTPUT 🦋

The json output format is intended to be both human readable and convenient for automated parsing. For the most part its sections mirror those of the normal output. The runtime value is reported in msec and the bw value is reported in 1024 bytes per second units. 🎀

🧚 JSON+ OUTPUT 🧚

The json+ output format is identical to the json output format, except that it adds a full dump of the completion latency bins. ✨ Each bins object contains a set of (key, value) pairs where keys are latency durations and values count how many I/Os had completion latencies of the corresponding duration. For example, consider:

"bins" : { "87552" : 1, "89600" : 1, "94720" : 1, "96768" : 1,
"97792" : 1, "99840" : 1, "100864" : 2, "103936" : 6, "104960" :
534, "105984" : 5995, "107008" : 7529, ... }

This data indicates that one I/O required 87,552ns to complete, two I/Os required 100,864ns to complete, and 7529 I/Os required 107,008ns to complete. Look at those numbers go, honestly obsessed. 🔥

Also included with fio is a Python script fio_jsonplus_clat2csv that takes json+ output and generates CSV-formatted latency data suitable for plotting. 📊

The latency durations actually represent the midpoints of latency intervals. For details refer to stat.h in the fio source. 👑

💅 TRACE FILE FORMAT 🎀

Okay so real talk: you're gonna run into two trace file formats out in the wild, and bestie, one of them is ancient. The older one (v1) has been unsupported since version 1.20-rc3 (March 2008) — like, that's older than some of us 😭 — but we're still gonna describe it down below just in case you stumble on a vintage trace and wanna know what you're looking at. History lesson, but make it fio.

Either way, the trace is just a plain text file, one single action per line. Simple. Clean. Aesthetic. ✨

Trace file format v1

Every line is one I/O action, like this:

rw, offset, length

where rw=0/1 for read/write, and the offset and length entries are in bytes.

Heads up bestie: this format is not supported in fio versions >= 1.20-rc3. It's a museum piece. 💀

Trace file format v2

Version 2 dropped back in fio version 1.17, and it ate 💅 — it lets you access more than one file per trace and gives you a way bigger set of possible file actions.

The first line of the trace file has to be:

"fio version 2 iolog"

After that you can have lines in two different formats, described below. Let's gooo.

The file management format:

filename action

The filename is given as an absolute path. The action can be one of these:

add — add the given filename to the trace.
open — open the file with the given filename. The filename has to have been added with the add action before. (Order matters, queen 👑)
close — close the file with the given filename. The file has to have been opened before.

The file I/O action format:

filename action offset length

The filename is given as an absolute path, and has to have been added and opened before it can be used with this format. The offset and length are given in bytes. The action can be one of these:

wait — wait for offset microseconds. Everything below 100 is discarded. The time is relative to the previous wait statement. Note that action wait is not allowed as of version 3, since the same behavior can be achieved using timestamps.
read — read length bytes beginning from offset.
write — write length bytes beginning from offset.
sync — fsync(2) the file.
datasync — fdatasync(2) the file.
trim — trim the given file from the given offset for length bytes.

Trace file format v3

Version 3 showed up in fio version 3.31, and it's giving accountability — it forces each action to have a timestamp associated with it. No more vibes-based ordering, we're being precise now. 🦋

The first line of the trace file has to be:

"fio version 3 iolog"

Following this can be lines in two different formats, which are described below.

The file management format:

timestamp filename action

The file I/O action format:

timestamp filename action offset length

The timestamp is relative to the beginning of the run (ie starts at 0). The filename, action, offset and length are identical to version 2, except that version 3 does not allow the wait action. Rip wait, you had a good run 🫶

🔥 I/O REPLAY - MERGING TRACES 🌈

Colocation is a super common practice for getting the most out of a machine — stack those workloads, bestie. But knowing which workloads play nicely with each other and which ones absolutely do not? That's the much harder task. 😭

While fio can replay workloads concurrently via multiple jobs, it leaves some variability up to the scheduler, which makes results harder to reproduce. Merging is the way to make the order of events consistent. Iconic behavior, honestly.

Merging is integrated into I/O replay and happens when a merge_blktrace_file is specified. The list of files passed to read_iolog go through the merge process and output a single file stored to the specified file. The output file is passed on as if it were the only file passed to read_iolog. An example would look like:

$ fio --read_iolog="<file1>:<file2>"
--merge_blktrace_file="<output_file>"

Creating only the merged file can be done by passing the command line argument merge-blktrace-only. Just the merge, no I/O, we love a minimalist. ✨

Scaling traces can be done to see the relative impact of any particular trace being slowed down or sped up. merge_blktrace_scalars takes in a colon separated list of percentage scalars. It is index paired with the files passed to read_iolog. Matchy matchy.

With scaling, it may be desirable to match the running time of all traces. This can be done with merge_blktrace_iters. It is index paired with read_iolog just like merge_blktrace_scalars.

In an example, given two traces, A and B, each 60s long. If we want to see the impact of trace A issuing IOs twice as fast and repeat trace A over the runtime of trace B, the following can be done:

$ fio --read_iolog="<trace_a>:"<trace_b>"
--merge_blktrace_file"<output_file>"
--merge_blktrace_scalars="50:100" --merge_blktrace_iters="2:1"

This runs trace A at 2x the speed twice for approximately the same runtime as a single run of trace B. Slay. 🏃‍♀️💨

🦋 CPU IDLENESS PROFILING ✨

Sometimes we wanna understand CPU overhead in a test. Like, for example, we're testing patches for the specific goodness of whether they reduce CPU usage. Fio implements a balloon approach to create a thread per CPU that runs at idle priority — meaning it only runs when nobody else needs the cpu. Polite thread. Respectful thread. 🎀 By measuring the amount of work completed by the thread, idleness of each CPU can be derived accordingly.

An unit work is defined as touching a full page of unsigned characters. Mean and standard deviation of time to complete an unit work is reported in "unit work" section. Options can be chosen to report detailed percpu idleness or overall system idleness by aggregating percpu stats. Numbers, but make it cute. 💖

👑 VERIFICATION AND TRIGGERS 💖

Fio is usually run in one of two ways when data verification is done. Let me lay it out for you, bestie:

A normal write job of some sort with verify enabled. When the write phase has completed, fio switches to reads and verifies everything it wrote. Classic.
Running just the write phase, and then later on running the same job (but with reads instead of writes) to repeat the same I/O patterns and verify the contents.

Both of these methods depend on the write phase being completed, because fio otherwise has no idea how much data was written. No notes. 💀

With verification triggers, fio supports dumping the current write state to local files. Then a subsequent read verify workload can load this state and know exactly where to stop. This is useful for testing cases where power is cut to a server in a managed fashion, for instance. Dramatic, but make it controlled. 🔥

A verification trigger consists of two things:

Storing the write state of each job.
Executing a trigger command.

The write state is relatively small, on the order of hundreds of bytes to single kilobytes. It contains information on the number of completions done, the last X completions, etc. Compact queen. ✨

A trigger is invoked either through creation (touch) of a specified file in the system, or through a timeout setting. If fio is run with --trigger-file=/tmp/trigger-file, then it will continually check for the existence of /tmp/trigger-file. When it sees this file, it will fire off the trigger (thus saving state, and executing the trigger command). It's basically "notice me" energy, but for files. 🙈

For client/server runs, there's both a local and remote trigger. If fio is running as a server backend, it will send the job states back to the client for safe storage, then execute the remote trigger, if specified. If a local trigger is specified, the server will still send back the write state, but the client will then execute the trigger. Teamwork makes the dream work. 🫶

Verification trigger example

Let's say we want to run a powercut test on the remote Linux machine 'server'. Our write workload is in write-test.fio. We want to cut power to 'server' at some point during the run, and we'll run this test from the safety or our local machine, 'localbox'. On the server, we'll start the fio backend normally:

server# fio --server

and on the client, we'll fire off the workload:

localbox$ fio --client=server
--trigger-file=/tmp/my-trigger --trigger-remote="bash -c
"echo b > /proc/sysrq-triger""

We set /tmp/my-trigger as the trigger file, and we tell fio to execute:

echo b > /proc/sysrq-trigger

on the server once it has received the trigger and sent us the write state. This will work, but it's not really cutting power to the server, it's merely abruptly rebooting it. If we have a remote way of cutting power to the server through IPMI or similar, we could do that through a local trigger command instead. Let's assume we have a script that does IPMI reboot of a given hostname, ipmi-reboot. On localbox, we could then have run fio with a local trigger instead:

localbox$ fio --client=server
--trigger-file=/tmp/my-trigger --trigger="ipmi-reboot
server"

For this case, fio would wait for the server to send us the write state, then execute ipmi-reboot server when that happened. Chef's kiss. 👨‍🍳💋

Loading verify state

To load stored write state, a read verification job file must contain the verify_state_load option. If that is set, fio will load the previously stored state. For a local fio run this is done by loading the files directly, and on a client/server run, the server backend will ask the client to send the files over and load them from there. And that's the whole loop, bestie — state saved, state loaded, nothing lost. 💜

🌸 LOG FILE FORMATS 🌸

Okay bestie, so fio keeps receipts — it can log your latencies, your bandwidth, and your IOPS in a whole bunch of log file formats. But they all share one common format, and it looks like this:

time (msec), value, data direction, block size (bytes), offset
(bytes), command priority, issue time (nsec)

The Time for a log entry is always in milliseconds, no exceptions. The value though? That depends on which type of log you're looking at, and it'll be one of these:

Latency log — the value is latency in nsecs 💅
Bandwidth log — the value is in KiB/sec 🚀
IOPS log — the value is IOPS ⚡

Data direction is one of these three, and it's giving very much numeric:

0 — the I/O is a READ
1 — the I/O is a WRITE
2 — the I/O is a TRIM

The entry's block size is always in bytes, periodt. The offset is the position in bytes from the start of the file for that particular I/O — and you can toggle whether it gets logged at all with log_offset 🎀

Now, the Command priority field, okay, listen up:

If log_prio is not set, the entry's Command priority is 1 for an I/O executed with the highest RT priority class (prioclass=1 or cmdprio_class=1) and 0 otherwise. This whole thing is controlled by the prioclass option and the ioengine-specific cmdprio_percentage / cmdprio_class options, iykyk.

If log_prio is set, then the entry's Command priority is the priority set for the I/O, expressed as a 16-bit hexadecimal number — with the lowest 13 bits indicating the priority value (that's the prio and cmdprio options) and the highest 3 bits indicating the IO priority class (that's prioclass and cmdprio_class). Math queen behavior, honestly 🔥

The entry's issue time is the command issue time in nanoseconds, and you can toggle whether it's logged with log_issue_time. This field has valid values in the completion latency log file (clat), or the submit latency log file (slat). In other log files the field just has value 0. Not every log gets to be special 💀

By default fio logs every single individual I/O — very main character, very thorough. But when windowed logging is set through log_avg_msec, you get either the average (that's the default), the maximum (log_window_value is set to max) value seen over the specified period of time, or both the average value and maximum value1 (log_window_value is set to both) recorded.

And when both values are reported, the log file format takes this form:

time (msec), value, value1, data direction, block size (bytes),
offset (bytes), command priority, issue time (nsec)

Each data direction seen within the window period will aggregate its values in a separate row. Also, heads up: when you're using windowed logging, the block size, offset and issue time entries will always contain 0. Every time. No cap. 🙈

🦋 CLIENT / SERVER 🦋

Normally you just run fio as a stand-alone application on the machine where the I/O workload should be generated. Simple, cute, done. However — the backend and frontend of fio can be run separately, meaning the fio server can generate an I/O workload on the "Device Under Test" while being controlled by a client on another machine. Long distance relationship but make it storage 💖

So start the server on the machine that has access to the storage DUT:

$ fio --server=args

Where args defines what fio listens to. The arguments are of the form type,hostname or IP,port. And type is either:

ip (or ip4) for TCP/IP v4
ip6 for TCP/IP v6
sock for a local unix domain socket

hostname is either a hostname or an IP address, and port is the port to listen to — which is only valid for TCP/IP, not a local socket, bestie, don't try it.

Some examples, because I know you love a demo:

fio --server — start a fio server, listening on all interfaces on the default port (8765).
fio --server=ip:hostname,4444 — start a fio server, listening on the IP belonging to hostname and on port 4444.
fio --server=ip6:::1,4444 — start a fio server, listening on IPv6 localhost ::1 and on port 4444.
fio --server=,4444 — start a fio server, listening on all interfaces on port 4444.
fio --server=1.2.3.4 — start a fio server, listening on IP 1.2.3.4 on the default port.
fio --server=sock:/tmp/fio.sock — start a fio server, listening on the local socket /tmp/fio.sock.

Once a server is running, a "client" can connect to the fio server with:

$ fio <local-args> --client=<server> <remote-args> <job file(s)>

Where local-args are arguments for the client where it is running, server is the connect string, and remote-args and job file(s) are sent to the server. The server string follows the same format as it does on the server side, so you can pass IP/hostname/socket and port strings. ✨

⚠️ Note that all job options must be defined in job files when running fio as a client. Any job options specified in remote-args will be ignored. Straight up ignored. Nothing you can do about it, bestie 😭

Fio can connect to multiple servers this way:

$ fio --client=<server1> <job file(s)> --client=<server2> <job
file(s)>

If the job file is located on the fio server, then you can tell the server to load a local file as well. This is done by using --remote-config:

$ fio --client=server --remote-config /path/to/file.fio

Then fio will open this local (to the server) job file instead of being passed one from the client. Convenient 🫶

If you have many servers (example: 100 VMs/containers — iconic, honestly), you can input a pathname of a file containing host IPs/names as the parameter value for the --client option. For example, here's an example host.list file containing 2 hostnames:

host1.your.dns.domain
host2.your.dns.domain

The fio command would then be:

$ fio --client=host.list <job file(s)>

💀 In this mode, you cannot input server-specific parameters or job files — all servers receive the same job file. Equality, I guess.

Now, in order to let fio --client runs use a shared filesystem from multiple hosts, fio --client now prepends the IP address of the server to the filename. For example, if fio is using the directory /mnt/nfs/fio and is writing filename fileio.tmp, with a --client hostfile containing two hostnames h1 and h2 with IP addresses 192.168.10.120 and 192.168.10.121, then fio will create two files:

/mnt/nfs/fio/192.168.10.120.fileio.tmp
/mnt/nfs/fio/192.168.10.121.fileio.tmp

This behavior can be disabled by the unique_filename option. Your files, your rules 💅

Terse output in client/server mode will differ slightly from what's produced when fio is run in stand-alone mode. See the terse output section for details — it's up there somewhere, you got this.

Also! If one fio invocation runs workloads on multiple servers, fio will provide at the end an aggregate summary report for all workloads. This aggregate summary report assumes that options affecting reporting like unified_rw_reporting and percentile_list are identical across all the jobs summarized. Having different values for these options is an unsupported configuration. So keep 'em matching, bestie 🙏

👑 AUTHORS 👑

fio was written by Jens Axboe <axboe@kernel.dk>.

This man page was written by Aaron Carroll <aaronc@cse.unsw.edu.au> based on documentation by Jens Axboe.

This man page was rewritten by Tomohiro Kusumi <tkusumi@tuxera.com> based on documentation by Jens Axboe.

Shoutout to the whole squad, fr fr 🫶

🎀 REPORTING BUGS 🎀

Report bugs to the fio mailing list <fio@vger.kernel.org>.

See REPORTING-BUGS.

REPORTING-BUGS: https://git.kernel.org/pub/scm/linux/kernel/git/axboe/fio.git/tree/REPORTING-BUGS

🌈 SEE ALSO 🌈

For further documentation see HOWTO and README.

Sample jobfiles are available in the examples/ directory. These are typically located under /usr/share/doc/fio.

HOWTO: https://git.kernel.org/pub/scm/linux/kernel/git/axboe/fio.git/tree/HOWTO

README: https://git.kernel.org/pub/scm/linux/kernel/git/axboe/fio.git/tree/README.rst

And that's a wrap on the whole fio(1) page, bestie!! 🎉 This was the User Manual, dated May 2025, and it's fio(1) — you made it all the way to the end, I'm so proud of you, go take a screenshot for the group chat 💖✨🦋

📦 where this page comes from

arch rolling: fio 3.42-1 · GPL2
nixpkgs unstable: fio 3.42
💖 girlified by deepseek-ai/DeepSeek-V4.1-Flash-1017 on 2026-11-04: a language model's rewrite of the original, under its terms