srun(1)

💅 srun(1) — run parallel jobs, bestie ✨

okay so this is srun, from the Slurm Commands manual, and honestly? it's giving main character energy. it's literally the command you use to run a parallel job on a cluster that Slurm manages 🎀 lowkey iconic.

---

🌸 SYNOPSIS 🌸

srun [OPTIONS(0)... [executable(0) [args(0)...]]] [ : [OPTIONS(N)...]]
executable(N) [args(N)...]

so the teeeeea is that the option(s) define multiple jobs in a co-scheduled heterogeneous job 🦋. if you wanna know more about heterogeneous jobs, the doc is right here: https://slurm.schedmd.com/heterogeneous_jobs.html — go read it, i'll wait 💜

---

✨ DESCRIPTION ✨

Run a parallel job on a cluster managed by Slurm. and ngl, if it needs to, srun will first create a resource allocation to run that parallel job in. slay, it handles it for you 💅

also, there's a whole document about how the various options influence the allocation of cpus to jobs and tasks — it's here, no cap: https://slurm.schedmd.com/cpu_management.html 📚

---

🫶 RETURN VALUE 🫶

srun will return the highest exit code of all tasks run, or the highest signal (with the high-order bit set in an 8-bit integer — like, e.g., 128 + signal) of any task that exited with a signal 🫡

⚠️ the value 253 is reserved for out-of-memory errors. yes, oom. that's a whole vibe and not a good one 😭

and when you invoke it with --async, srun returns 0 once the controller accepts the step-create request — regardless of how the step's tasks ultimately exit 🙈. to actually peek at the step's exit code after it completes, use sacct(1) ⭐

---

🧚 EXECUTABLE PATH RESOLUTION 🧚

the executable is resolved in this exact order, bestie:

1. If executable starts with ".", then path is constructed as: current working directory / executable
2. If executable starts with a "/", then path is considered absolute.
3. If executable can be resolved through PATH. See path_resolution(7).
4. If executable is in current working directory.

and about that current working directory — it's the calling process working directory unless the --chdir argument is passed, which will override the current working directory 💖

---

✨ OPTIONS ✨

• -A, --account=<account> — Charge resources used by this job to specified account. the account is an arbitrary string, fr fr. the account name may be changed after job submission using the scontrol command. This option applies to job allocations 💸
• --acctg-freq=<datatype>=<interval>[,<datatype>=<interval>...] — Define the job accounting and profiling sampling intervals in seconds ⏱️. This can be used to override the JobAcctGatherFrequency parameter in the slurm.conf file. <datatype>=<interval> specifies the task sampling interval for the jobacct_gather plugin or a sampling interval for a profiling type by the acct_gather_profile plugin. Multiple comma-separated <datatype>=<interval> pairs may be specified. supported datatype values are:
  • task — Sampling interval for the jobacct_gather plugins and for task profiling by the acct_gather_profile plugin. NOTE: This frequency is used to monitor memory usage. If memory limits are enforced the highest frequency a user can request is what is configured in the slurm.conf file. It can not be disabled. 🧠
  • energy — Sampling interval for energy profiling using the acct_gather_energy plugin. ⚡
  • network — Sampling interval for infiniband profiling using the acct_gather_interconnect plugin. 🌐
  • filesystem — Sampling interval for filesystem profiling using the acct_gather_filesystem plugin. 🍓

the default value for the task sampling interval is 30 seconds. the default value for all other intervals is 0. An interval of 0 disables sampling of the specified type. If the task sampling interval is 0, accounting information is collected only at job termination (reducing Slurm interference with the job). Smaller (non-zero) values have a greater impact upon job performance, but a value of 30 seconds is not likely to be noticeable for applications having less than 10,000 tasks. This option applies to job allocations 💅

• --async — Submit a job step that runs asynchronously, instead of synchronously with the srun command that submitted it 🔥. The step is launched and supervised by the job's step manager, which will queue the step for later execution if resources are not currently available. The srun command returns immediately after the step is accepted by the controller and reports its step ID, without waiting for the step to start or finish. See --parsable for details on output format 👀

Output is written directly on the compute nodes to the files named by --output and --error; if neither is given, slurm-%J.out (relative to each task's working directory) is used. Input is also supported by naming a file with --input=<file>. Forms that require srun to forward its own stdin to the tasks (--input=all or --input=<taskid>) do not work because srun has no I/O connection to the step 😭. Some options, such as --label and --unbuffered, are applied on the compute node by the slurmstepd(8) process that writes the output, rather than by srun.

Since srun exits before the step finishes, it does not propagate the step's exit code and does not forward signals (e.g., SIGINT from Ctrl+C) to the running tasks 🫡. Use scancel <jobid>.<stepid> to signal or terminate an asynchronous step.

This option can only be used in an existing job allocation that has the step manager enabled (see --stepmgr in salloc(1)/sbatch(1) and SlurmctldParameters=enable_stepmgr in slurm.conf(5)).

This option is mutually exclusive with --pty, --immediate, and any MPI type other than none, whether selected with --mpi, the SLURM_MPI_TYPE environment variable, or MpiDefault in slurm.conf(5); use --mpi=none to override a non-none default 🚫

⚠️ NOTE: This is a new feature that is still in beta, its usage may change in future releases. don't say i didn't warn you 💀

• --bb=<spec> — Burst buffer specification 🎀. The form of the specification is system dependent. Also see --bbf. This option applies to job allocations. When the --bb option is used, Slurm parses this option and creates a temporary burst buffer script file that is used internally by the burst buffer plugins. See Slurm's burst buffer guide for more information and examples: https://slurm.schedmd.com/burst_buffer.html ✨
• --bbf=<file_name> — Path of file containing burst buffer specification. The form of the specification is system dependent. Also see --bb. This option applies to job allocations. See Slurm's burst buffer guide for more information and examples: https://slurm.schedmd.com/burst_buffer.html 📂
• --bcast[=<dest_path>] — Copy executable file to allocated compute nodes 🚀. If a file name is specified, copy the executable to the specified destination file path. If the path specified ends with '/' it is treated as a target directory, and the destination file name will be slurm_bcast_<job_id>.<step_id>_<nodename>. If no dest_path is specified and the slurm.conf BcastParameters DestDir is configured then it is used, and the filename follows the above pattern. If none of the previous is specified, then --chdir is used, and the filename follows the above pattern too. For example, "srun --bcast=/tmp/mine -N3 a.out" will copy the file "a.out" from your current directory to the file "/tmp/mine" on each of the three allocated compute nodes and execute that file. This option applies to step allocations 💅
• --bcast-exclude={NONE|<exclude_path>[,<exclude_path>...]} — Comma-separated list of absolute directory paths to be excluded when autodetecting and broadcasting executable shared object dependencies through --bcast. If the keyword "NONE" is configured, no directory paths will be excluded. The default value is that of slurm.conf BcastExclude and this option overrides it. See also --bcast and --send-libs 🦋
• -b, --begin=<time> — Defer initiation of this job until the specified time ⏰. It accepts times of the form HH:MM:SS to run a job at a specific time of day (seconds are optional). (If that time is already past, the next day is assumed.) You may also specify midnight, noon, elevenses (11 AM), fika (3 PM) or teatime (4 PM) and you can have a time-of-day suffixed with AM or PM for running in the morning or the evening 🌅. You can also say what day the job will be run, by specifying a date of the form MMDDYY or MM/DD/YY YYYY-MM-DD. Combine date and time using the following format YYYY-MM-DD[THH:MM[:SS]]. You can also give times like now + count time-units, where the time-units can be seconds (default), minutes, hours, days, or weeks. The keywords today and tomorrow may also be used. The value may be changed after job submission using the scontrol command 🫶. For example:

--begin=16:00
--begin=now+1hour
--begin=now+60           (seconds by default)
--begin=2010-01-20T12:34:00

Notes on date/time specifications, bestie 📝:

• Although the 'seconds' field of the HH:MM:SS time specification is allowed by the code, note that the poll time of the Slurm scheduler is not precise enough to guarantee dispatch of the job on the exact second 😭. The job will be eligible to start on the next poll following the specified time. The exact poll interval depends on the Slurm scheduler (e.g., 60 seconds with the default sched/builtin).
• If no time (HH:MM:SS) is specified, the default is (00:00:00).
• If a date is specified without a year (e.g., MM/DD) then the current year is assumed, unless the combination of MM/DD and HH:MM:SS has already passed for that year, in which case the next year is used.

This option applies to job allocations 💜

• -D, --chdir=<path> — Have the remote processes do a chdir to path before beginning execution 📁. The default is to chdir to the current working directory of the srun process. The path can be specified as full path or relative path to the directory where the command is executed. This option applies to job allocations ✨
• --cluster-constraint=<list> — Specifies features that a federated cluster must have to have a sibling job submitted to it 🌈. Slurm will attempt to submit a sibling job to a cluster if it has at least one of the specified features.

---

more parts coming, don't go anywhere 💖

• -M, --clusters=<string> — which cluster(s) you're firing commands at 💫. Multiple cluster names? Just comma-separate them, easy. The job gets submitted to whichever one gives the earliest expected job initiation time (yes, that's the flex — fastest start wins). Default value is the current cluster. If you pass all, it'll query to run on every cluster, main character behaviour. Peep the --export option for controlling environment variables exported between clusters. Quick heads-up: this option applies only to job allocations. And ngl, the slurmdbd has to be up for this to work properly, unless you're running in a federation with FederationParameters=fed_display configured. Slay.
• --comment=<string> — literally just an arbitrary comment. Yours, free-form, no notes. This option applies to job allocations 🎀.
• --compress[=type] — squishes the file before sending it over to the compute hosts 🗜️✨. The optional argument tells it which data compression library to use; the default is BcastParameters Compression= if that's set, or "lz4" otherwise. Supported values are "lz4" — that's the list, bestie, it's short. Some compression libraries may be unavailable on some systems (rude but true). This one's for use with the --bcast option, and it applies to step allocations.
• --consolidate-segments — makes sure all the segments from the allocation get consolidated into one higher-level aggregated block 🧩.

This option applies to job allocations. NOTE: it will only work with the topology/block plugin. No plugin, no consolidation, no exceptions 💀.

• -C, --constraint=<list> — okay so the Slurm admin can assign features to nodes, and this is where you say which of those features your job absolutely needs 👑. If you're hunting for 'soft' constraints, that's a different thing — go see --prefer. Only nodes packing features that match your job's constraints will be used to satisfy the request. You can stack multiple constraints with AND, OR, matching OR, resource counts, etc. — though lowkey some operators aren't supported on all system types 🙈.

NOTE: Changeable features are features defined by a NodeFeatures plugin. That's the whole definition, bookmark it.

Supported --constraint options include:

Single Name

Only nodes which have the specified feature will be used. For example, --constraint="intel". Simple, iconic, gets the job done 💅.

Node Count

A request can specify the number of nodes needed with some feature by appending an asterisk and count after the feature name. For example, --nodes=16 --constraint="graphics*4" indicates that the job requires 16 nodes and that at least four of those nodes must have the feature "graphics". If requesting more than one feature and using node counts, the request must have square brackets surrounding it 🫶.

NOTE: This option is not supported by the helpers NodeFeatures plugin. Heterogeneous jobs can be used instead — workaround era.

AND

Only nodes with all of the specified features will be used. The ampersand is the AND operator, and yes it's giving 2009 HTML. For example, --constraint="intel&gpu" 🔥.

OR

Only nodes with at least one of the specified features will be used. The vertical bar is the OR operator. If changeable features are not requested, nodes in the allocation can have different features. For example, salloc -N2 --constraint="intel|amd" can result in a job allocation where one node has the intel feature and the other node has the amd feature 🌈.

However, if the expression contains a changeable feature, then all OR operators are automatically treated as Matching OR so that all nodes in the job allocation have the same set of features. For example, salloc -N2 --constraint="foo|bar&baz" — the job is allocated two nodes where both nodes have foo, or bar and baz (one or both nodes could have foo, bar, and baz). The helpers NodeFeatures plugin will find the first set of node features that matches all nodes in the job allocation; these features are set as active features on the node and passed to RebootProgram (see slurm.conf(5)) and the helper script (see helpers.conf(5)). In this case, the helpers plugin uses the first of "foo" or "bar,baz" that match the two nodes in the job allocation. Whew 😮‍💨.

Matching OR

If only one of a set of possible options should be used for all allocated nodes, then use the OR operator and enclose the options within square brackets. For example, --constraint="[rack1|rack2|rack3|rack4]" might be used to specify that all nodes must be allocated on a single rack of the cluster, but any of those four racks can be used. One rack to rule them all 🎀.

Multiple Counts

Specific counts of multiple resources may be specified by using the AND operator and enclosing the options within square brackets. For example, --constraint="[rack1*2&rack2*4]" might be used to specify that two nodes must be allocated from nodes with the feature of "rack1" and four nodes must be allocated from nodes with the feature "rack2" 📊.

NOTE: This option is not supported by the helpers NodeFeatures plugin.

NOTE: Multiple Counts can cause jobs to be allocated with a non-optimal network layout. Consider yourself warned 💀.

Brackets

Brackets can be used to indicate that you are looking for a set of nodes with the different requirements contained within the brackets. For example, --constraint="[(rack1|rack2)*1&(rack3)*2]" will get you one node with either the "rack1" or "rack2" features and two nodes with the "rack3" feature 🧚. If requesting more than one feature and using node counts, the request must have square brackets surrounding it.

NOTE: Brackets are only reserved for Multiple Counts and Matching OR syntax. AND operators require a count for each feature inside square brackets (i.e. "[quad*2&hemi*1]"). Slurm will only allow a single set of bracketed constraints per job 🙈.

NOTE: Square brackets are not supported by the helpers NodeFeatures plugin. Matching OR can be requested without square brackets by using the vertical bar character with at least one changeable feature. loophole unlocked ✨

Parentheses

Parentheses can be used to group like node features together. For example, --constraint="[(knl&snc4&flat)*4&haswell*1]" might be used to specify that four nodes with the features "knl", "snc4" and "flat" plus one node with the feature "haswell" are required 🫶.

Parentheses can also be used to group operations. Without parentheses, node features are parsed strictly from left to right. For example, --constraint="foo&bar|baz" requests nodes with foo and bar, or baz. --constraint="foo|bar&baz" requests nodes with foo and baz, or bar and baz (note how baz was AND'd with everything — she's involved). --constraint="foo&(bar|baz)" requests nodes with foo and at least one of bar or baz. NOTE: OR within parentheses should not be used with a KNL NodeFeatures plugin but is supported by the helpers NodeFeatures plugin ✨.

WARNING: When srun is executed from within salloc or sbatch, the constraint value can only contain a single feature name. None of the other operators are currently supported for job steps 😭. This option applies to job and step allocations.

• --container=<path_to_container> — absolute path to the OCI container bundle. Absolute, bestie, don't come at me with relative paths 🐳.
• --container-id=<container_id> — unique name for the OCI container. Make it cute, make it unique 💅.
• --container-type=<container_type> — the job container type for the job. That's it, that's the whole thing 🎀.
• --contiguous — if set, then the allocated nodes must form a contiguous set 🧩✨.

NOTE: whether this option is honored depends on the topology plugin. It is enforced by topology/flat and topology/tree. Under topology/tree it also disables the topology-aware node selection optimizations, so the resulting allocation may span nodes with greater network distance between them 😬. It is silently ignored by all other topology plugins — no error, no warning, just vibes and ghosts. This option applies to job allocations.

• -S, --core-spec=<num> — count of Specialized Cores per node reserved by the job for system operations and not used by the application 🔒. If AllowSpecResourcesUsage is enabled, a job can override the CoreSpecCount of all its allocated nodes with this option. The overridden Specialized Cores will still be reserved for system processes. The job will get an implicit --exclusive allocation for the rest of the Cores on the nodes, resulting in the job's processes being able to use (and being charged for 💸) all the Cores on the nodes except for the overridden Specialized Cores. This option can not be used with the --thread-spec option. pick one, queen.

NOTE: Explicitly setting a job's specialized core value implicitly sets the --exclusive option. Sneaky.

NOTE: This option may implicitly impact the number of tasks if -n was not specified. so like... heads up.

This option applies to job allocations.

• --cores-per-socket=<cores> — restricts node selection to nodes with at least the specified number of cores per socket 🧠. See the additional info under the -B option above when the task/affinity plugin is enabled. This option applies to job allocations.
• --cpu-bind=[{quiet|verbose},]<type> — binds tasks to CPUs 🔗, and it's used only when the task/affinity plugin is enabled. NOTE: to have Slurm always report on the selected CPU binding for all commands executed in a shell, you can enable verbose mode by setting the SLURM_CPU_BIND environment variable value to "verbose". The way that's so handy.

The following informational environment variables are set when --cpu-bind is in use:

SLURM_CPU_BIND_VERBOSE
SLURM_CPU_BIND_TYPE
SLURM_CPU_BIND_LIST

✨

Okay so real talk about CPU binding, bestie, because this is where it gets spicy 🌶️

Peep the ENVIRONMENT VARIABLES section for the full tea on each individual SLURM_CPU_BIND variable. Tiny catch though: those variables are available only if the task/affinity plugin is configured. No plugin, no variables, no vibes 💅

When you're using --cpus-per-task to run multithreaded tasks, heads up: CPU binding is inherited from the parent of the process. Translation? Your multithreaded task should either specify or clear the CPU binding itself, or else every single thread is gonna camp out on the same mask/CPU as the parent 😭 Alternatively, you could use fat masks (masks which specify more than one allowed CPU) for the tasks so there's multiple CPUs available for the multithreaded tasks. Do that and you're golden ✨

Also ngl, a job step can be allocated different numbers of CPUs on each node, or be allocated CPUs not starting at location zero 🙈 Which is exactly why one of the options which automatically generate the task binding is the recommended move. Explicitly specified masks or bindings are only honored when the job step has been allocated every available CPU on the node. Periodt.

Binding a task to a NUMA locality domain means binding the task to the set of CPUs that belong to the NUMA locality domain or "NUMA node". And if NUMA locality domain options are used on systems with no NUMA support? Then each socket is considered a locality domain. We improvise 💫

If the --cpu-bind option is not used, the default binding mode depends upon Slurm's configuration and the step's resource allocation, in this exact order:

1. If all allocated nodes have the same configured CpuBind mode, that will be used.
2. Otherwise if the job's Partition has a configured CpuBind mode, that will be used.
3. Otherwise if Slurm has a configured TaskPluginParam value, that mode will be used.
4. Otherwise automatic binding will be performed as described below 👇

🌀 Auto Binding 🌀

Applies only when task/affinity is enabled. If the job step allocation includes an allocation with a number of sockets, cores, or threads equal to the number of tasks times cpus-per-task, then the tasks will by default be bound to the appropriate resources (auto binding). Iconic. Slay. 🫶

Want it off? Disable this mode of operation by explicitly setting "--cpu-bind=none". And use TaskPluginParam=autobind=[threads|cores|sockets] to set a default cpu binding in case "auto binding" doesn't find a match.

🎛️ Supported options include: 🎛️

• q[uiet] — quietly bind before task runs (this is the default, lowkey)
• v[erbose] — verbosely report binding before task runs (for the girlies who need the receipts 👀)
• no[ne] — do not bind tasks to CPUs (default unless auto binding is applied)
• map_cpu:<list> — bind by setting CPU masks on tasks (or ranks) as specified, where <list> is <cpu_id_for_task_0>,<cpu_id_for_task_1>,... If the number of tasks (or ranks) exceeds the number of elements in this list, elements in the list will be reused as needed starting from the beginning of the list. To simplify support for large task counts, the lists may follow a map with an asterisk and repetition count. For example "map_cpu:0*4,3*4".
• mask_cpu:<list> — bind by setting CPU masks on tasks (or ranks) as specified, where <list> is <cpu_mask_for_task_0>,<cpu_mask_for_task_1>,... The mapping is specified for a node and identical mapping is applied to the tasks on every node (i.e. the lowest task ID on each node is mapped to the first mask specified in the list, etc.). CPU masks are always interpreted as hexadecimal values but can be preceded with an optional '0x'. If the number of tasks (or ranks) exceeds the number of elements in this list, elements in the list will be reused as needed starting from the beginning of the list. To simplify support for large task counts, the lists may follow a map with an asterisk and repetition count. For example "mask_cpu:0x0f*4,0xf0*4".
• rank_ldom — bind to a NUMA locality domain by rank. Not supported unless the entire node is allocated to the job 💀
• map_ldom:<list> — bind by mapping NUMA locality domain IDs to tasks as specified, where <list> is <ldom1>,<ldom2>,...<ldomN>. The locality domain IDs are interpreted as decimal values unless they are preceded with '0x', in which case they are interpreted as hexadecimal values. Not supported unless the entire node is allocated to the job 💀
• mask_ldom:<list> — bind by setting NUMA locality domain masks on tasks as specified, where <list> is <mask1>,<mask2>,...<maskN>. NUMA locality domain masks are always interpreted as hexadecimal values but can be preceded with an optional '0x'. Not supported unless the entire node is allocated to the job 💀
• sockets — automatically generate masks binding tasks to sockets. Only the CPUs on the socket which have been allocated to the job will be used. If the number of tasks differs from the number of allocated sockets this can result in sub-optimal binding 😬
• cores — automatically generate masks binding tasks to cores. If the number of tasks differs from the number of allocated cores this can result in sub-optimal binding 😬
• threads — automatically generate masks binding tasks to threads. If the number of tasks differs from the number of allocated threads this can result in sub-optimal binding 😬
• ldoms — automatically generate masks binding tasks to NUMA locality domains. If the number of tasks differs from the number of allocated locality domains this can result in sub-optimal binding 😬
• help — show help message for cpu-bind 🆘

This option applies to job and step allocations.

---

⚡ --cpu-freq=<p1>[-p2][:p3] ⚡

Request that the job step initiated by this srun command be run at some requested frequency if possible, on the CPUs selected for the step on the compute node(s). Basically: "can u go faster pls" but formally 🌸

p1 can be [#### | low | medium | high | highm1] which will set the frequency scaling_speed to the corresponding value, and set the frequency scaling_governor to UserSpace. See below for definition of the values.

p1 can also be [Conservative | OnDemand | Performance | PowerSave] which will set the scaling_governor to the corresponding value. The governor has to be in the list set by the slurm.conf option CpuFreqGovernors.

When p2 is present, p1 will be the minimum scaling frequency and p2 will be the maximum scaling frequency. In that case the governor p3 or CpuFreqDef cannot be UserSpace since it doesn't support a range. Sorry bestie 💔

p2 can be [#### | medium | high | highm1]. p2 must be greater than p1 and is incompatible with UserSpace governor.

p3 can be [Conservative | OnDemand | Performance | PowerSave | SchedUtil | UserSpace] which will set the governor to the corresponding value.

If p3 is UserSpace, the frequency scaling_speed, scaling_max_freq and scaling_min_freq will be statically set to the value defined by p1.

Any requested frequency below the minimum available frequency will be rounded to the minimum available frequency. Same energy in reverse: any requested frequency above the maximum available frequency will be rounded to the maximum available frequency. The hardware said "no ❤️" and did its own thing.

The CpuFreqDef parameter in slurm.conf will be used to set the governor in absence of p3. If there's no CpuFreqDef, the default governor will be to use the system current governor set in each cpu. Specifying a range without CpuFreqDef or a specific governor is therefore not allowed 🚫

Acceptable values at present include:

• #### — frequency in kilohertz
• Low — the lowest available frequency
• High — the highest available frequency
• HighM1 — (high minus one) will select the next highest available frequency
• Medium — attempts to set a frequency in the middle of the available range
• Conservative — attempts to use the Conservative CPU governor
• OnDemand — attempts to use the OnDemand CPU governor (the default value)
• Performance — attempts to use the Performance CPU governor
• PowerSave — attempts to use the PowerSave CPU governor
• UserSpace — attempts to use the UserSpace CPU governor

The following informational environment variable is set in the job step when --cpu-freq option is requested:

SLURM_CPU_FREQ_REQ

This environment variable can also be used to supply the value for the CPU frequency request if it is set when the 'srun' command is issued. The --cpu-freq on the command line will override the environment variable value. The form on the environment variable is the same as the command line. See the ENVIRONMENT VARIABLES section for a description of the SLURM_CPU_FREQ_REQ variable 💖

NOTE: this one's a request, not a demand, bestie. If the job step's node can't do CPU frequency setting, or the value you asked for is outside the bounds of the legal frequencies, it logs an error but the job step keeps going. No dramatic shutdown 💅

NOTE: setting the frequency for just the CPUs of the job step implies the tasks are confined to those CPUs. If task confinement isn't configured (i.e. the task/affinity TaskPlugin is enabled, or the task/cgroup TaskPlugin is enabled with "ConstrainCores=yes" set in cgroup.conf), this parameter is straight-up ignored. 🙈

NOTE: when the step completes, the frequency and governor of each selected CPU gets reset to the previous values. We clean up after ourselves, periodt. ✨

NOTE: when submitting jobs with the --cpu-freq option with linuxproc as the ProctrackType, jobs can run too quickly before Accounting is able to poll for job information. Result? Not all of the accounting info will be present 😭

This option applies to job and step allocations.

• --cpus-per-gpu=<ncpus> — request that ncpus processors be allocated per allocated GPU. This option implies --exact. Not compatible with the --cpus-per-task option. 🫶

This option applies to job and step allocations.

• -c, --cpus-per-task=<ncpus> — request that ncpus be allocated per process. Highkey useful if your job is multithreaded and needs more than one CPU per task for optimal performance. Explicitly requesting this option implies --exact. The default is one CPU per process and does not imply --exact. If -c is specified without -n, as many tasks will be allocated per node as possible while satisfying the -c restriction. For instance, on a cluster with 8 CPUs per node, a job request for 4 nodes and 3 CPUs per task may be allocated 3 or 6 CPUs per node (1 or 2 tasks per node) depending upon resource consumption by other jobs. Such a job may be unable to execute more than a total of 4 tasks. 💀

WARNING: there are configurations and options interpreted differently by job and job step requests which can result in inconsistencies for this option. For example, srun -c2 --threads-per-core=1 prog may allocate two cores for the job, but if each of those cores contains two threads, the job allocation will include four CPUs. The job step allocation will then launch two threads per CPU for a total of two tasks. 😵‍💫

WARNING: when srun is executed from within salloc or sbatch, there are configurations and options which can result in inconsistent allocations when -c has a value greater than -c on salloc or sbatch.

NOTE: if --mem-per-cpu is also specified, the number of allocated cpus can be increased if MaxMemPerCPU is exceeded. In the case -n is not specified, the number of tasks can be higher than expected. 👀

This option applies to job and step allocations.

• --deadline=<OPT> — remove the job if no ending is possible before this deadline (start > (deadline - time[-min])). Default is no deadline. Note that if neither DefaultTime nor MaxTime are configured on the partition the job is in, the job will need to specify some form of time limit (--time[-min]) if a deadline is to be used. 🕐

Valid time formats are:

HH:MM[:SS] [AM|PM]
MMDD[YY] or MM/DD[/YY] or MM.DD[.YY]
MM/DD[/YY]-HH:MM[:SS]
YYYY-MM-DD[THH:MM[:SS]]]
now[+count[seconds(default)|minutes|hours|days|weeks]]

This option applies only to job allocations.

• --delay-boot=<minutes> — do not reboot nodes in order to satisfy this job's feature specification if the job has been eligible to run for less than this time period. If the job has waited for less than the specified period, it will use only nodes which already have the specified features. The argument is in units of minutes. A default value may be set by a system administrator using the delay_boot option of the SchedulerParameters configuration parameter in the slurm.conf file, otherwise the default value is zero (no delay). 🦋

This option applies only to job allocations.

• -d, --dependency=<dependency_list> — defer the start of this job until the specified dependencies have been satisfied. Once a dependency is satisfied, it is removed from the job. This option does not apply to job steps (executions of srun within an existing salloc or sbatch allocation) only to job allocations. <dependency_list> is of the form <type:job_id[:job_id][,type:job_id[:job_id]]> or <type:job_id[:job_id][?type:job_id[:job_id]]>. All dependencies must be satisfied if the "," separator is used. Any dependency may be satisfied if the "?" separator is used. Only one separator may be used. For instance:

-d afterok:20:21,afterany:23

means that the job can run only after a 0 return code of jobs 20 and 21 AND the completion of job 23. However:

-d afterok:20:21?afterany:23

means that any of the conditions (afterok:20 OR afterok:21 OR afterany:23) will be enough to release the job. Many jobs can share the same dependency and these jobs may even belong to different users. The value may be changed after job submission using the scontrol command. Dependencies on remote jobs are allowed in a federation. Once a job dependency fails due to the termination state of a preceding job, the dependent job will never be run, even if the preceding job is requeued and has a different termination state in a subsequent execution. This option applies to job allocations. 👑

• after:job_id[[+time][:jobid[+time]...]] — after the specified jobs start or are cancelled and time in minutes from job start or cancellation happens, this job can begin execution. If no time is given then there is no delay after start or cancellation.
• afterany:job_id[:jobid...] — this job can begin execution after the specified jobs have terminated. This is the default dependency type. ✨
• afterburstbuffer:job_id[:jobid...] — this job can begin execution after the specified jobs have terminated and any associated burst buffer stage out operations have completed.
• aftercorr:job_id[:jobid...] — a task of this job array can begin execution after the corresponding task ID in the specified job has completed successfully (ran to completion with an exit code of zero). If the specified job is not an array, this is treated the same as afterok.
• afternotok:job_id[:jobid...] — this job can begin execution after the specified jobs have terminated in some failed state (non-zero exit code, node failure, timed out, etc). This job must be submitted while the specified job is still active or within MinJobAge seconds after the specified job has ended. If the dependent job ID is not found and is on the same cluster as the job submission, the job is rejected. If the dependent job ID is not found and is on a different cluster from the job submission, the dependency is marked as failed. 😭
• afterok:job_id[:jobid...] — this job can begin execution after the specified jobs have successfully executed (ran to completion with an exit code of zero). This job must be submitted while the specified job is still active or within MinJobAge seconds after the specified job has ended. If the dependent job ID is not found and is on the same cluster as the job submission, the job is rejected. If the dependent job ID is not found and is on a different cluster from the job submission, the dependency is marked as failed.
• singleton — this job can begin execution after any previously launched jobs sharing the same job name and user have terminated. In other words, only one job by that name and owned by that user can be running or suspended at any point in time. In a federation, a singleton dependency must be fulfilled on all clusters unless DependencyParameters=disable_remote_singleton is used in slurm.conf. 💅

• -X, --disable-status — disable the display of task status when srun receives a single SIGINT (Ctrl+C). Instead immediately forward the SIGINT to the running job. Without this option a second Ctrl+C in one second is required to forcibly terminate the job and srun will immediately exit. May also be set via the environment variable SLURM_DISABLE_STATUS. This option applies to job allocations. 🔥
• -m, --distribution={*|block|cyclic|arbitrary|plane=<size>}[:{*|block|cyclic|fcyclic}[:{*|block|cyclic|fcyclic}]][,{Pack|NoPack}] — specify alternate distribution methods for remote processes. For job allocation, this sets environment variables that will be used by subsequent srun requests. Task distribution affects job allocation at the last stage of the evaluation of available resources by the cons_tres plugin. Consequently, other options (e.g. --ntasks-per-node, --cpus-per-task) may affect resource selection prior to task distribution. To ensure a specific task distribution, jobs should have access to entire nodes, which can be accomplished by using the --exclusive flag or by requesting all the resources on the node(s). 🌈

okay so buckle up bestie, because this next bit is the whole distribution system and honestly it's giving "three-layer cake of task placement." 🎂 The way this works: the option string has up to three distribution methods glued together with : colons, and each layer does a different job. Let me break it down so your brain doesn't short-circuit:

The first distribution method (everything before the first :) controls how tasks get spread across nodes. The second one (after the first :) controls how allocated CPUs get spread across sockets for binding to tasks. The third one (after the second :) controls how allocated CPUs get spread across cores for binding to tasks. The second and third distributions only do anything if task affinity is enabled 💅, and the third one is only supported if the task/cgroup plugin is configured. The default value for each distribution type is specified by *. Slay, a wildcard.

Heads up: with select/cons_tres, the number of CPUs allocated to each socket and node may be different. Not all nodes are built the same, bestie 😭. There's more info on resource allocation, distribution of tasks to nodes, and binding of tasks to CPUs at https://slurm.schedmd.com/mc_support.html.

First distribution method (distribution of tasks across nodes):

• * — use the default method for distributing tasks to nodes (block). But plot twist 👀: if a topology plugin is configured and you didn't specify a distribution method, tasks get distributed following topology order rather than node index order. See CR_NO_DIST_TOPO_BLOCK in SelectTypeParameters to disable this behavior.
• block — tasks get distributed so that consecutive tasks share a node. Example time: imagine three nodes, each with two cpus. A four-task block distribution request puts tasks one and two on the first node, task three on the second node, and task four on the third node. block is also the default behavior if the number of tasks exceeds the number of allocated nodes. Iconic, honestly.
• cyclic — tasks get distributed so consecutive tasks go over consecutive nodes (round-robin style 🔄). Same example: three nodes, each with two cpus, four-task cyclic request → tasks one and four on the first node, task two on the second node, task three on the third node. Note that when SelectType is select/cons_tres, the same number of CPUs may not be allocated on each node — task distribution will be round-robin among all the nodes with CPUs yet to be assigned to tasks. cyclic is the default behavior if the number of tasks is no larger than the number of allocated nodes.
• plane — tasks get distributed in blocks of size <size>. The size must be given, or SLURM_DIST_PLANESIZE must be set. The number of tasks distributed to each node is the same as for cyclic, but the taskids assigned to each node depend on the plane size. Additional distribution specifications cannot be combined with this option. For more details (including examples and diagrams — cute little diagrams 🎀), please see https://slurm.schedmd.com/mc_support.html and https://slurm.schedmd.com/dist_plane.html
• arbitrary — allocates processes in-order as listed in the file designated by the environment variable SLURM_HOSTFILE. If this variable is listed it will override any other method specified. If it's not set, the method defaults to block. The hostfile must contain at minimum the number of hosts requested, and be one per line or comma separated. If you specify a task count (-n, --ntasks=<number>), your tasks will be laid out on the nodes in the order of the file.

⚠️ NOTE: the arbitrary distribution option is supported only with the generic allocation method and cannot be used with features such as topology, CR_LLN, CR_Pack_Nodes, or pack_serial_at_end. This option is meant primarily to control a job step's task layout in an existing job allocation for the srun command.

⚠️ NOTE, part two, electric boogaloo: if the number of tasks is given and a list of requested nodes is also given, the number of nodes used from that list will be reduced to match the number of tasks if the number of nodes in the list is greater than the number of tasks. Basically it trims the list to fit ✂️.

Second distribution method (distribution of CPUs across sockets for binding):

• * — use the default method for distributing CPUs across sockets (cyclic).
• block — distributes allocated CPUs consecutively from the same socket for binding to tasks, before using the next consecutive socket.
• cyclic — distributes allocated CPUs for binding to a given task consecutively from the same socket, and from the next consecutive socket for the next task, round-robin across sockets. Tasks requiring more than one CPU will have all of those CPUs allocated on a single socket if possible.

⚠️ NOTE: on nodes with hyper-threading enabled, a task not requesting full cores may be distributed across sockets 😬. You can avoid this by specifying --ntasks-per-core=1, which forces tasks to allocate full cores. Fix it with one flag, bestie.

• fcyclic — distributes allocated CPUs for binding to tasks from consecutive sockets in a round-robin fashion across the sockets. Tasks requiring more than one CPU will have each CPU allocated in a cyclic fashion across sockets.

Third distribution method (distribution of CPUs across cores for binding):

• * — use the default method for distributing CPUs across cores (inherited from the second distribution method 👑).
• block — distributes allocated CPUs consecutively from the same core for binding to tasks, before using the next consecutive core.
• cyclic — distributes allocated CPUs for binding to a given task consecutively from the same core, and from the next consecutive core for the next task, round-robin across cores.
• fcyclic — distributes allocated CPUs for binding to tasks from consecutive cores in a round-robin fashion across the cores.

Optional control for task distribution over nodes:

• Pack — instead of evenly distributing a job step's tasks across its allocated nodes, pack them as tightly as possible on the nodes 🫂. This only applies when the block task distribution method is used.
• NoPack — instead of packing a job step's tasks as tightly as possible on the nodes, distribute them evenly. This user option will supersede the SelectTypeParameters CR_Pack_Nodes configuration parameter. User says no packing, so no packing 💅.

This option applies to job and step allocations.

---

✨ --epilog={none|<executable>} ✨

srun will run executable just after the job step completes. The command line arguments for executable will be the command and arguments of the job step. If none is specified, then no srun epilog will be run. This parameter overrides the SrunEpilog parameter in slurm.conf. And listen — this parameter is completely independent from the Epilog parameter in slurm.conf. Different things, don't mix them up 🙈. This option applies to job allocations.

🌸 -e, --error=<filename_pattern> 🌸

Tells srun how stderr is supposed to be redirected. By default, in interactive mode, srun redirects stderr to the same file as stdout, if one is specified. The --error option exists so you can send stdout and stderr to different locations, which is honestly so valid of it. See IO Redirection below for more options. If the specified file already exists, it will be overwritten — no mercy, no backup 😭. This option applies to job and step allocations.

🦋 --exact 🦋

Lets a step access only the resources requested for the step. By default, all non-GRES resources on each node in the step allocation will be used. This option only applies to step allocations.

⚠️ NOTE: parallel steps will either be blocked or rejected until the requested step resources are available, unless --overlap is specified. Also, job resources can be held after the completion of an srun command while Slurm does job cleanup, and step epilogs and/or SPANK plugins can further delay the release of step resources. Patience is a virtue, bestie ⏳.

💜 -x, --exclude={<host1[,<host2>...]|<filename>} 💜

Requests that a specific list of hosts not be included in the resources allocated to this job 🚫. The host list will be assumed to be a filename if it contains a / character. Sneaky but fair. This option applies to job and step allocations.

👑 --exclusive[={allocation|mcs|topo|user}] 👑

Requests exclusive resources, either for the job or the job step depending on context. When srun --exclusive is used to initiate a step within a running job, it's a synonym for --exact. Same energy, different name 💅.

Okay so back to --exclusive for a sec, because this is the main character of resource-sharing rules 💅. When srun --exlusive is used to initiate a job and a step, BOTH whole nodes AND --exact behavior will be executed — meaning all resources of a node(s) will be allocated to the job, and only specific resources will be granted to each task within the step. The allocation option will get whole nodes but NOT apply the --exact behavior — meaning each step will get access to all resources within the allocation (similar to running srun without other limiting options in a running allocation). The allocation option can be made default by setting LaunchParameters=srun_exclusive_allocation in slurm.conf. ✨

Options (mcs|topo|user) exist solely to control how multiple jobs share resources when srun is used to allocate a job and do not imply --exact. When used to initiate a job, the job allocation can not share nodes (or topology segment with the "=topo") with other running jobs (or just other users with the "=user" option or "=mcs" option). If user/mcs/topo are not specified (i.e. the job allocation can not share nodes with other running jobs), the job allocation is allocated all CPUs and GRES on all nodes in the allocation, but is only allocated as much memory as it requested. This is by design to support gang scheduling, because suspended jobs still reside in memory. To request all the memory on a node, use --mem=0. The default shared/exclusive behavior depends on system configuration and the partition's OverSubscribe option takes precedence over the job's option.

NOTE: Since shared GRES (MPS) cannot be allocated at the same time as a sharing GRES (GPU) this option only allocates all sharing GRES and no underlying shared GRES.

NOTE: This option is mutually exclusive with --oversubscribe.

See EXAMPLE below. 👑

---

• --export={[ALL,]<environment_variables>|ALL|NONE} — Identify which environment variables from the submission environment are propagated to the launched application. 🫶
  • --export=ALL — Default mode if --export is not specified. All of the user's environment will be loaded from the caller's environment.
  • --export=NONE — None of the user environment will be defined. User must use absolute path to the binary to be executed that will define the environment. User can not specify explicit environment variables with "NONE". This option is particularly important for jobs that are submitted on one cluster and execute on a different cluster (e.g. with different paths). To avoid steps inheriting environment export settings (e.g. "NONE") from sbatch command, either set --export=ALL or the environment variable SLURM_EXPORT_ENV should be set to "ALL".
  • --export=[ALL,]<environment_variables> — Exports all SLURM* environment variables along with explicitly defined variables. Multiple environment variable names should be comma separated. Environment variable names may be specified to propagate the current value (e.g. "--export=EDITOR") or specific values may be exported (e.g. "--export=EDITOR=/bin/emacs"). If "ALL" is specified, then all user environment variables will be loaded and will take precedence over any explicitly given environment variables.
  • Example: --export=EDITOR,ARG1=test — In this example, the propagated environment will only contain the variable EDITOR from the user's environment, SLURM_* environment variables, and ARG1=test.
  • Example: --export=ALL,EDITOR=/bin/emacs — There are two possible outcomes for this example. If the caller has the EDITOR environment variable defined, then the job's environment will inherit the variable from the caller's environment. If the caller doesn't have an environment variable defined for EDITOR, then the job's environment will use the value given by --export.
• --external-launcher — Create a special step on one or more allocated nodes which won't consume any resources, but will have access to all of the job's allocated resources on the nodes. 🧚 Options like --ntasks-per-*, --mem*, --cpus*, --tres*, --gres*, will be ignored. This is meant for use MPI implementations that require their own launcher. This launches a step with access to all the resources and which will later spawn any number of user processes with access to all these resources. The resource usage within this special step will still be accounted for if the accounting plugins are enabled. This special step can be overlapped with any other step.

NOTE: This option is not intended to be used directly.

• --extra=<string> — An arbitrary string enclosed in single or double quotes if using spaces or some special characters. If SchedulerParameters=extra_constraints is enabled, this string is used for node filtering based on the Extra field in each node.
• -B, --extra-node-info=<sockets>[:cores[:threads]] — Restrict node selection to nodes with at least the specified number of sockets, cores per socket and/or threads per core. NOTE: These options do not specify the resource allocation size. Each value specified is considered a minimum. An asterisk (*) can be used as a placeholder without restricting node selection for that type. The individual levels can also be specified in separate options if desired:
  • --sockets-per-node=<sockets>
  • --cores-per-socket=<cores>
  • --threads-per-core=<threads>

If task/affinity plugin is enabled, then specifying an allocation in this manner also sets a default --cpu-bind option of threads if the -B option specifies a thread count, otherwise an option of cores if a core count is specified, otherwise an option of sockets. If SelectType is configured to select/cons_tres, it must have a parameter of CR_Core, CR_Core_Memory, CR_Socket, or CR_Socket_Memory for this option to be honored. If not specified, the scontrol show job will display 'ReqS:C:T=*:*:*'. This option applies to job allocations.

NOTE: This option is mutually exclusive with --hint, --threads-per-core and --ntasks-per-core.

NOTE: If the number of sockets, cores and threads were all specified, the number of nodes was specified (as a fixed number, not a range) and the number of tasks was NOT specified, srun will implicitly calculate the number of tasks as one task per thread.

• --gpu-bind=[verbose,]<type> — Equivalent to --tres-bind=gres/gpu:[verbose,]<type> See --tres-bind for all options and documentation.
• --gpu-freq=[<type]=value>[,<type=value>][,verbose] — Request that GPUs allocated to the job are configured with specific frequency values. This option can be used to independently configure the GPU and its memory frequencies. After the job is completed, the frequencies of all affected GPUs will be reset to the highest possible values. In some cases, system power caps may override the requested values. The field type can be "memory". If type is not specified, the GPU frequency is implied. The value field can either be "low", "medium", "high", "highm1" or a numeric value in megahertz (MHz). If the specified numeric value is not possible, a value as close as possible will be used. See below for definition of the values. The verbose option causes current GPU frequency information to be logged. Examples of use include "--gpu-freq=medium,memory=high" and "--gpu-freq=450". 🌈

Supported value definitions:

• low — the lowest available frequency.
• medium — attempts to set a frequency in the middle of the available range.
• high — the highest available frequency.
• highm1 — (high minus one) will select the next highest available frequency.

• -G, --gpus=[type:]<number> — Specify the total number of GPUs required for the job. An optional GPU type specification can be supplied. See also the --gpus-per-node, --gpus-per-socket and --gpus-per-task options. NOTE: The allocation has to contain at least one GPU per node, or one of each GPU type per node if types are used. Use heterogeneous jobs if different nodes need different GPU types.
• --gpus-per-node=[type:]<number> — Specify the number of GPUs required for the job on each node included in the job's resource allocation. An optional GPU type specification can be supplied. For example "--gpus-per-node=volta:3". Multiple options can be requested in a comma separated list, for example: "--gpus-per-node=volta:3,kepler:1". See also the --gpus, --gpus-per-socket and --gpus-per-task options.

NOTE: This option is mutually exclusive with --gres=gpu.

• --gpus-per-socket=[type:]<number> — Specify the number of GPUs required for the job on each socket included in the job's resource allocation. An optional GPU type specification can be supplied. For example "--gpus-per-socket=volta:3". Multiple options can be requested in a comma separated list, for example: "--gpus-per-socket=volta:3,kepler:1". Requires job to specify a sockets per node count (--sockets-per-node). See also the --gpus, --gpus-per-node and --gpus-per-task options. This option applies to job allocations. 💜
• --gpus-per-task=[type:]<number> — how many GPUs each task needs within the job's resource allocation 💅 you can optionally say which kind of GPU, like --gpus-per-task=volta:1. Wanna mix and match? Go off — comma-separated lists are valid: --gpus-per-task=volta:3,kepler:1. Also peep --gpus, --gpus-per-socket and --gpus-per-node while you're here. Heads up bestie: this one needs an explicit task count — so -n, --ntasks, or --gpus=X --gpus-per-task=Y — not a fuzzy range of nodes via -N, --nodes (ambiguous node ranges are so not the vibe here). This option will implicitly set --tres-bind=gres/gpu:per_task:<gpus_per_task>, or if you asked for multiple GPU types it sets --tres-bind=gres/gpu:per_task:<gpus_per_task_type_sum>. But you can totally override that with an explicit --tres-bind=gres/gpu spec. slay 🔥
• --gres=<list> — a comma-delimited list of generic consumable resources you want per node 🎀 every entry follows the format name[[:type]:count]:
  • name is the type of consumable resource (e.g. gpu)
  • type is an optional classification of the resource (e.g. a100)
  • count is how many of those you want — default value is 1 The count can carry a suffix: k/K (multiple of 1024), m/M (multiple of 1024 x 1024), g/G (multiple of 1024 x 1024 x 1024), t/T (multiple of 1024 x 1024 x 1024 x 1024), p/P (multiple of 1024 x 1024 x 1024 x 1024 x 1024). The resources you list get allocated to the job on each node. What's actually available is configurable by your system administrator, so it's giving ✨customizable✨. If the option argument is help, srun prints the list of available generic consumable resources and then exits. Iconic examples: --gres=gpu:2, --gres=gpu:kepler:2, and --gres=help.

NOTE: this option applies to job and step allocations. By default, a job step is allocated all of the generic resources that have been requested by the job — except those implicitly requested when a job is exclusive. Wanna flip that so each job step gets no generic resources? You have three routes: explicitly set the value of --gres to specify zero counts for each generic resource, OR set --gres=none, OR set the SLURM_STEP_GRES environment variable to none. pick your fighter 💀

NOTE: this option is mutually exclusive with --gpus-per-node. literally cannot do both 🙈

• --gres-flags=<type> — pick your generic resource task binding options, bestie:
  • allow-task-sharing — lets tasks access every GPU in the job's allocation that lives on the same node as the task. Super useful when you're using --gpu-bind or --tres-bind=gres/gpu to bind GPUs to specific tasks, but you still want the tasks gossiping with each other over GPU comms 🫶

NOTE: this option is specific to srun.

• multiple-tasks-per-sharing — negates one-task-per-sharing. useful if it's already set by default in SelectTypeParameters.
• disable-binding — negates enforce-binding. useful if it's already set by default in SelectTypeParameters.
• enforce-binding — the only CPUs available to the job will be the ones bound to the selected GRES (i.e. the CPUs identified in the gres.conf file get strictly enforced). Fair warning: this can delay the job's start 🕰️ for example, a job needing two GPUs and one CPU will be delayed until both GPUs on a single socket are available, rather than using GPUs bound to separate sockets — but the app's performance may improve thanks to better communication speed. Requires the node to be configured with more than one socket, and resource filtering happens on a per-socket basis.

NOTE: this option can be set by default in SelectTypeParameters. NOTE: this option is specific to SelectType=cons_tres for job allocations. NOTE: this option can give undefined results if you try to enforce binding on multiple gres across multiple sockets. seriously, careful 😭

• one-task-per-sharing — don't let different tasks be allocated shared gres from the same sharing gres. no sharing the sharing gres, periodt.

NOTE: this flag is only enforced if shared gres are requested with --tres-per-task. NOTE: this option can be set by default with SelectTypeParameters=ONE_TASK_PER_SHARING_GRES. NOTE: this option is specific to SelectTypeParameters=MULTIPLE_SHARING_GRES_PJ.

• -h, --help — shows you the help info and dips ✌️
• --het-group=<expr> — tells srun which component(s) of a heterogeneous job allocation you want a step created for. Only applies to srun commands issued inside an salloc allocation or an sbatch script. expr is a set of integers matching one or more option offsets on the salloc or sbatch command line. Examples: --het-group=2, --het-group=0,4, --het-group=1,3-5. The default value is --het-group=0. so lowkey chill by default 💜
• --hint=<type> — bind tasks according to what your app is hinting it wants ✨

NOTE: this option implies specific values for certain related options, which totally blocks it from being used with any user-specified values for --ntasks-per-core, --cores-per-socket, --sockets-per-node, --threads-per-core, --cpu-bind (other than --cpu-bind=verbose) or -B. Those conflicting options will override --hint when specified as command line arguments. But if a conflicting option is set as an environment variable, then --hint as a command line argument takes precedence. iykyk 👑

• compute_bound — settings for compute-bound apps: use all cores in each socket, one thread per core.
• memory_bound — settings for memory-bound apps: use only one core in each socket, one thread per core.
• multithread — use extra threads with in-core multi-threading, which can benefit communication-intensive apps. Only supported with the task/affinity plugin.
• nomultithread — don't use extra threads with in-core multi-threading; restricts tasks to one thread per core. Only supported with the task/affinity plugin.
• help — show this help message.

This option applies to job allocations.

• -H, --hold — submit the job in a held state (priority of zero) 🧊 a held job can be released later with scontrol to reset its priority, e.g. scontrol release <job_id>. This option applies to job allocations.
• --ignore-signals=<signals_list> — a comma-separated list of signals for srun to straight-up ignore 🙈 both signal names and numbers are accepted. Normally srun reacts to these signals by forwarding them to the running tasks or by performing cleanup — but when a signal is listed here, srun takes no action on it. Ignoring signals may interfere with mechanisms for cancelling or otherwise controlling job behavior, so carefully consider the details below before using this option. Valid signals and their default handling are as follows:
  • SIGINT — triggers an interactive abort sequence. Ignoring this will prevent srun termination with Ctrl+C.
  • SIGQUIT — forwarded as SIGKILL.
  • SIGTERM — forwarded as SIGKILL. Ignoring this inside sbatch/salloc allocations may delay job termination upon scancel or timeout.
  • SIGHUP — forwarded as SIGKILL. Ignoring this may delay job termination on terminal hangup or at the end of an interactive salloc allocation, and interfere with cleanup tasks.
  • SIGUSR1 — forwarded as-is to tasks.
  • SIGUSR2 — forwarded as-is to tasks.

lowkey this one is a "know what you're doing" moment, bestie 🫶

• -I, --immediate[=<seconds>] — exit if resources aren't available within the time period you specify ⏰ if no argument is given (seconds defaults to 1), resources must be available immediately for the request to succeed. If defer is configured in SchedulerParameters and seconds=1, the allocation request will fail immediately; defer conflicts with and takes precedence over this option. By default --immediate is off, and the command will just block until resources become available (main character energy: it waits 💅). Since this option's argument is optional, for proper parsing the single-letter option must be followed immediately by the value with no space between them — so -I60 and not -I 60. This option applies to job and step allocations.
• -i, --input=<mode> — how should stdin be redirected? By default, srun redirects stdin from the terminal to all tasks. Peep "IO Redirection" below for more options. On OS X, the poll() function doesn't support stdin, so input from a terminal isn't possible 🍎😭 This option applies to job and step allocations.
• -J, --job-name=<jobname> — name your job, bestie 🎀 the specified name shows up alongside the job ID number when you query running jobs on the system. The default is the supplied executable program's name.

NOTE: this information may be written to the slurm_jobacct.log file. That file is space delimited, so if you put a space in the jobname it will cause problems properly displaying the contents of slurm_jobacct.log when the sacct command is used. no spaces, i'm begging 🙏

This option applies to job and step allocations.

• --jobid=<jobid> — kick off a job step under an already-allocated job with job ID id 🦋 using this option makes srun behave exactly as if the SLURM_JOB_ID environment variable was set. This option applies to step allocations.
• -K, --kill-on-bad-exit[=0|1] — the "bestie, if one of your tasks flops, do we nuke the whole step?" switch 💀. If you don't pass it, the default comes from the Slurm config parameter KillOnBadExit. If you do pass it, it overrides KillOnBadExit, periodt. An option argument of zero means don't terminate the job; a non-zero argument or no argument at all means do terminate the job. Heads up: this option takes precedence over the -W, --wait option, so it'll kill the job immediately if a task exits with a non-zero exit code. And because the argument is optional, for proper parsing you have to glue the value right onto the single letter with no space between them — so -K1, not -K 1 🫶.
• -l, --label — prepends the task number to lines of stdout/err ✨. The --label option will prepend lines of output with the remote task id. This one applies to step allocations.
• -L, --licenses=<license>[@db][:count][,license[@db][:count]...] — tell Slurm which licenses (or other resources available on all nodes of the cluster) have to be allocated to this job 🎀. License names can be followed by a colon and a count (the default count is one). You can request multiple licenses. If they're separated by a comma (, meaning AND), then all the requested licenses are required for the job — for example, --licenses=foo:4,bar. If they're separated by a pipe (| meaning OR), then only one of the license requests is required for the job — for example, --licenses=foo:4|bar. AND and OR cannot both be used 🙅. To submit jobs using remote licenses (those served by slurmdbd), specify the name of the server providing the licenses — for example, --license=nastran@slurmdb:12. If LicenseParameters=RemoteFuzzyMatch is set in slurm.conf, it's possible to omit the server name.

NOTE: when submitting heterogeneous jobs, license requests may only be made on the first component job. For example srun -L ansys:2 : myexecutable.

NOTE: if licenses are tracked in AccountingStorageTres and OR is used, ReqTRES will display all requested tres separated by commas. AllocTRES will display only the license that was allocated to the job.

NOTE: when a job requests OR'd licenses, Slurm will attempt to allocate the licenses in the order in which they are requested. This specified order will take precedence even if the rest of the requested licenses could be satisfied on a requested reservation. This also applies to backfill planning when SchedulerParameters=bf_licenses is configured.

• --mail-type=<type> — get an email when certain event types occur 📬✨. The valid type values are NONE, BEGIN, END, FAIL, REQUEUE, ALL (equivalent to BEGIN, END, FAIL, INVALID_DEPEND, REQUEUE, and STAGE_OUT), INVALID_DEPEND (dependency never satisfied), STAGE_OUT (burst buffer stage out and teardown completed), TIME_LIMIT, TIME_LIMIT_90 (reached 90 percent of time limit), TIME_LIMIT_80 (reached 80 percent of time limit), and TIME_LIMIT_50 (reached 50 percent of time limit). Multiple type values may be specified in a comma separated list. NONE will suppress all event notifications, ignoring any other values specified. By default no email notifications are sent. The user to be notified is indicated with --mail-user. This option applies to job allocations 🫶.
• --mail-user=<user> — who gets those emails defined by --mail-type 💌. This may be a full email address or a username. If a username is specified, the value from MailDomain in slurm.conf will be appended to create an email address. The default value is the submitting user. This option applies to job allocations.
• --mcs-label=<mcs> — only used when a compatible MCSPlugin is enabled 🧚. This parameter is a group that the user belongs to (mcs/group) or an arbitrary label string (mcs/label). In both cases, no label will be assigned by default. This option applies to job allocations. Refer to the MCS documentation for more details: https://slurm.schedmd.com/mcs.html
• --mem=<size>[units] — the real memory required per node, no cap 🧠💅. Default units are mebibytes. Different units can be specified using the suffix [K|M|G|T]. The default value is DefMemPerNode and the maximum value is MaxMemPerNode. If configured, both of those parameters can be seen using the scontrol show config command. This parameter would generally be used if whole nodes are allocated to jobs (SelectType=select/linear). Specifying a memory limit of zero for a job step will restrict the job step to the amount of memory allocated to the job, but not remove any of the job's memory allocation from being available to other job steps. Also see --mem-per-cpu and --mem-per-gpu. The --mem, --mem-per-cpu and --mem-per-gpu options are mutually exclusive 🔥. If --mem, --mem-per-cpu or --mem-per-gpu are specified as command line arguments, then they will take precedence over the environment (potentially inherited from salloc or sbatch).

NOTE: a memory size specification of zero is treated as a special case and grants the job access to all of the memory on each node for newly submitted jobs and all available job memory to new job steps.

NOTE: the memory used by each slurmstepd process is included in the job's total memory usage. It typically consumes between 20MiB and 200MiB, though this can vary depending on system configuration and any loaded plugins.

NOTE: memory requests will not be strictly enforced unless Slurm is configured to use an enforcement mechanism. See ConstrainRAMSpace in the cgroup.conf(5) man page and OverMemoryKill in the slurm.conf(5) man page for more details.

This option applies to job and step allocations.

• --mem-bind=[{quiet|verbose},]<type> — bind your tasks to memory 🌸. Used only when the task/affinity plugin is enabled and the NUMA memory functions are available. Note that the resolution of CPU and memory binding may differ on some architectures — for example, CPU binding may be performed at the level of the cores within a processor while memory binding will be performed at the level of nodes, where the definition of "nodes" may differ from system to system. By default no memory binding is performed; any task using any CPU can use any memory. This option is typically used to ensure that each task is bound to the memory closest to its assigned CPU. The use of any type other than "none" or "local" is not recommended. If you want greater control, try running a simple test code with the options --cpu-bind=verbose,none --mem-bind=verbose,none to determine the specific configuration.

NOTE: to have Slurm always report on the selected memory binding for all commands executed in a shell, you can enable verbose mode by setting the SLURM_MEM_BIND environment variable value to "verbose".

The following informational environment variables are set when --mem-bind is in use:

SLURM_MEM_BIND_LIST
SLURM_MEM_BIND_PREFER
SLURM_MEM_BIND_TYPE
SLURM_MEM_BIND_VERBOSE

See the ENVIRONMENT VARIABLES section for a more detailed description of the individual SLURM_MEM_BIND* variables.

Supported options include:

• help — show this help message
• local — use memory local to the processor in use
• map_mem:<list> — bind by setting memory masks on tasks (or ranks) as specified, where <list> is <numa_id_for_task_0>,<numa_id_for_task_1>,... 🗺️. The mapping is specified for a node and identical mapping is applied to the tasks on every node (i.e. the lowest task ID on each node is mapped to the first ID specified in the list, etc.). NUMA IDs are interpreted as decimal values unless they are preceded with '0x', in which case they're interpreted as hexadecimal values. If the number of tasks (or ranks) exceeds the number of elements in this list, elements in the list will be reused as needed starting from the beginning of the list. To simplify support for large task counts, the lists may follow a map with an asterisk and repetition count. For example "map_mem:0x0f*4,0xf0*4". For predictable binding results, all CPUs for each node in the job should be allocated to the job.
• mask_mem:<list> — bind by setting memory masks on tasks (or ranks) as specified, where <list> is <numa_mask_for_task_0>,<numa_mask_for_task_1>,... 🎭. The mapping is specified for a node and identical mapping is applied to the tasks on every node (i.e. the lowest task ID on each node is mapped to the first mask specified in the list, etc.). NUMA masks are always interpreted as hexadecimal values. Note that masks must be preceded with a '0x' if they don't begin with [0-9] so they are seen as numerical values. If the number of tasks (or ranks) exceeds the number of elements in this list, elements in the list will be reused as needed starting from the beginning of the list. To simplify support for large task counts, the lists may follow a mask with an asterisk and repetition count. For example "mask_mem:0*4,1*4". For predictable binding results, all CPUs for each node in the job should be allocated to the job.
• no[ne] — don't bind tasks to memory (default) 🙈
• p[refer] — prefer use of first specified NUMA node, but permit use of other available NUMA nodes.
• q[uiet] — quietly bind before task runs (default) 🤫
• rank — bind by task rank (not recommended) 💀
• v[erbose] — verbosely report binding before task runs 📢

This option applies to job and step allocations ✨.

• --mem-per-cpu=<size>[units] — the minimum memory you need for every usable allocated CPU 🧠✨. Default units are mebibytes (yes, mebibytes, not megabytes, bestie, they're different and Slurm knows it). You can pick a different unit with the suffix [K|M|G|T]. The default value is DefMemPerCPU and the max is MaxMemPerCPU (there's an exception below, keep reading 📖). If both are configured you can peep them with the scontrol show config command.

Okay so here's the tea on what happens if your job's --mem-per-cpu value exceeds the configured MaxMemPerCPU: your limit gets treated as a memory limit per task instead. --mem-per-cpu gets reduced to a value no larger than MaxMemPerCPU, --cpus-per-task gets set, and the value of --cpus-per-task multiplied by the new --mem-per-cpu value will equal the original --mem-per-cpu value you specified. Math ate and left no crumbs 🧮💅.

This parameter would generally be used if individual processors are allocated to jobs (SelectType=select/cons_tres). But if resources are allocated by core, socket, or whole nodes, then the number of CPUs allocated to a job may be higher than the task count, and the value of --mem-per-cpu should be adjusted accordingly. Specifying a memory limit of zero for a job step will restrict the job step to the amount of memory allocated to the job, but not remove any of the job's memory allocation from being available to other job steps. Also see --mem and --mem-per-gpu. And listen: --mem, --mem-per-cpu and --mem-per-gpu are mutually exclusive — pick one, bestie, they do not share 🚫.

NOTE: If the final amount of memory requested by a job can't be satisfied by any of the nodes configured in the partition, the job will be rejected. This could happen if --mem-per-cpu is used with the --exclusive option for a job allocation and --mem-per-cpu times the number of CPUs on a node is greater than the total memory of that node. 💀

NOTE: This applies to usable allocated CPUs in a job allocation. This is important when more than one thread per core is configured. If a job requests --threads-per-core with fewer threads on a core than exist on the core (or --hint=nomultithread, which implies --threads-per-core=1), the job will be unable to use those extra threads on the core, and those threads will not be included in the memory per CPU calculation. But if the job has access to all threads on the core, those threads will be included in the memory per CPU calculation even if the job did not explicitly request those threads. 👀

In the following examples, each core has two threads.

In this first example, two tasks can run on separate hyperthreads in the same core because --threads-per-core is not used. The third task uses both threads of the second core. The allocated memory per cpu includes all threads:

$ salloc -n3 --mem-per-cpu=100
salloc: Granted job allocation 17199
$ sacct -j $SLURM_JOB_ID -X -o jobid%7,reqtres%35,alloctres%35
  JobID                             ReqTRES                           AllocTRES
------- ----------------------------------- -----------------------------------
  17199     billing=3,cpu=3,mem=300M,node=1     billing=4,cpu=4,mem=400M,node=1

In this second example, because of --threads-per-core=1, each task is allocated an entire core but is only able to use one thread per core. Allocated CPUs includes all threads on each core. However, allocated memory per cpu includes only the usable thread in each core:

$ salloc -n3 --mem-per-cpu=100 --threads-per-core=1
salloc: Granted job allocation 17200
$ sacct -j $SLURM_JOB_ID -X -o jobid%7,reqtres%35,alloctres%35
  JobID                             ReqTRES                           AllocTRES
------- ----------------------------------- -----------------------------------
  17200     billing=3,cpu=3,mem=300M,node=1     billing=6,cpu=6,mem=300M,node=1

• --mem-per-gpu=<size>[units] — the minimum memory required for every allocated GPU 🎮💖. Default units are mebibytes, and you can switch it up with the suffix [K|M|G|T]. Default value is DefMemPerGPU, available on both a global and per-partition basis. If configured, you can see the parameters with the scontrol show config and scontrol show partition commands. Also see --mem. And remember, --mem, --mem-per-cpu and --mem-per-gpu are mutually exclusive — one or the other, not all three, that's not the vibe 🌸.
• --mem-update=<margin>@<delay> — okay this one is lowkey iconic: it automatically reduces the job's memory limit after the job has been running for the time specified by delay 📉✨. The new limit is set to the job's current memory usage plus margin percent. The margin value is a percentage (e.g. 20 means 20% above current usage). The delay specifies how long after job start to sample memory usage and apply the reduction. Acceptable time formats for delay include "minutes", "minutes:seconds", "hours:minutes:seconds", "days-hours", "days-hours:minutes" and "days-hours:minutes:seconds".

The memory limit can only be reduced, never increased ⬇️. If the calculated new limit is not lower than the current limit, no change is made. A floor of 10% of the original memory request is enforced to prevent reducing the limit to an unreasonably small value. The memory usage is measured as RSS (resident physical memory) only — swap usage is not included. When ConstrainSwapSpace is configured, the swap limit is also reduced proportionally with the memory limit.

This option works with both --mem and --mem-per-cpu jobs. For --mem-per-cpu jobs, the new limit is applied as a per-node value, compared against the highest per-node allocation across all nodes. This is a one-shot operation: the reduction is applied at most once per job run. On job requeue, the auto-reduction may trigger again. This option applies to job allocations. It requires SlurmctldParameters=enable_stepmgr and JobAcctGatherType=jobacct_gather/cgroup to be configured 🫶.

• --mincpus=<n> — set a minimum number of logical cpus/processors per node 🖥️. This option applies to job allocations.
• --mpi=<mpi_type> — tells Slurm what kind of MPI you're rolling with. May result in unique initiation procedures, iykyk 🔥. The types:
  • cray_shasta — enable Cray PMI support. This is for applications built with the Cray Programming Environment. The PMI Control Port can be specified with the --resv-ports option or with the MpiParams=ports=<port range> parameter in your slurm.conf. This plugin does not have support for heterogeneous jobs. Support for cray_shasta is included by default ✅.
  • list — lists available mpi types to choose from 📋.
  • pmi2 — enable PMI2 support. The PMI2 support in Slurm works only if the MPI implementation supports it — in other words, if the MPI has the PMI2 interface implemented. --mpi=pmi2 will load the library lib/slurm/mpi_pmi2.so, which provides the server side functionality, but the client side must implement PMI2_Init() and the other interface calls.
  • pmix — enable PMIx support (https://pmix.github.io). The PMIx support in Slurm can be used to launch parallel applications (e.g. MPI) if it supports PMIx, PMI2 or PMI1. Slurm must be configured with pmix support by passing "--with-pmix=<PMIx installation path>" option to its "./configure" script.

At the time of writing, PMIx is supported in Open MPI starting from version 2.0. PMIx also supports backward compatibility with PMI1 and PMI2 and can be used if MPI was configured with PMI2/PMI1 support pointing to the PMIx library ("libpmix"). If MPI supports PMI1/PMI2 but doesn't provide the way to point to a specific implementation, a hack'ish solution leveraging LD_PRELOAD can be used to force "libpmix" usage 🙈.

• none — no special MPI processing. This is the default and works with many other versions of MPI 🎀.

This option applies to step allocations.

• --msg-timeout=<seconds> — tweak the job launch message timeout ⏱️. The default value is MessageTimeout in the Slurm configuration file slurm.conf. Changes to this are typically not recommended, but could be useful to diagnose problems. This option applies to job allocations.
• --multi-prog — run a job with different programs and different arguments for each task 🎭✨. In this case, the executable program specified is actually a configuration file specifying the executable and arguments for each task. See MULTIPLE PROGRAM CONFIGURATION below for details on the configuration file contents. This option applies to step allocations.
• --network=<type> — specify information pertaining to the switch or network 🌐. The interpretation of type is system dependent.

The network option is available on systems with HPE Slingshot networks. It can be used to request a job VNI (to be used for communication between job steps in a job). It also can be used to override the default network resources allocated for the job step. Multiple values may be specified in a comma-separated list.

• tcs=<class1>[:<class2>]... — set of traffic classes to configure for applications. Supported traffic classes are DEDICATED_ACCESS, LOW_LATENCY, BULK_DATA, and BEST_EFFORT. The traffic classes may also be specified as TC_DEDICATED_ACCESS, TC_LOW_LATENCY, TC_BULK_DATA, and TC_BEST_EFFORT. This option applies to the job allocation, but not to step allocations 👑.
• no_vni — don't allocate any VNIs for this job (even if multi-node) 🚫.
• job_vni — allocate a job VNI for this job.
• single_node_vni — allocate a job VNI for this job, even if it is a single-node job 🌱.
• adjust_limits — if set, slurmd will set an upper bound on network resource reservations by taking the per-NIC maximum resource quantity and subtracting the reserved or used values (whichever is higher) for any system network services; this is the default 💅.

• no_adjust_limits — okay so if this is set, slurmd works out its network resource reservations using only the per-resource configuration default and how many tasks the application has. It will not put an upper bound on those reservation requests based on how much the already-existing system network services are using. Heads up: that means more application launches could fail from network resource exhaustion 😭 — but if your app absolutely needs a certain amount of resources to function, this option makes sure it gets them. Periodt.
  • disable_rdzv_get — disables rendezvous gets in Slingshot NICs, which can improve performance for certain applications. ✨
  • nic_distribution_count=<val> — the number of NICs the user will evenly distribute their tasks over. Defaults to the number of NICs on each node.
  • def_<rsrc>=<val> — per-CPU reserved allocation for this resource.
  • res_<rsrc>=<val> — per-node reserved allocation for this resource. If set, it overrides the per-CPU allocation.
  • max_<rsrc>=<val> — maximum per-node limit for this resource.
  • depth=<depth> — multiplier for per-CPU resource allocation. Default is the number of reserved CPUs on the node.

The resources that may be requested are, and yes bestie there are a bunch 👑:

• txqs — transmit command queues. The default is 2 per-CPU, maximum 1024 per-node.
• tgqs — target command queues. The default is 1 per-CPU, maximum 512 per-node.
• eqs — event queues. The default is 2 per-CPU, maximum 2047 per-node.
• cts — counters. The default is 1 per-CPU, maximum 2047 per-node.
• tles — trigger list entries. The default is 1 per-CPU, maximum 2048 per-node.
• ptes — portable table entries. The default is 6 per-CPU, maximum 2048 per-node.
• les — list entries. The default is 16 per-CPU, maximum 16384 per-node.
• acs — addressing contexts. The default is 2 per-CPU, maximum 1022 per-node.

On systems configured with SwitchType=switch/nvidia_imex, the following options are supported:

• unique-channel-per-segment — instead of one channel for the entire job, allocate one channel per segment in the job. This only takes effect when topology/block is configured. 🦋

This option applies to job and step allocations.

• --nice[=adjustment] — run the job with an adjusted scheduling priority within Slurm. With no adjustment value, the scheduling priority is decreased by 100. A negative nice value increases the priority, otherwise it decreases it. The adjustment range is +/- 2147483645. Only privileged users can specify a negative adjustment. 💅
• -Z, --no-allocate — run the specified tasks on a set of nodes without creating a Slurm "job" in the Slurm queue structure, bypassing the normal resource allocation step entirely. The list of nodes must be given with the -w, --nodelist option. This is a privileged option, only available to the users "SlurmUser" and "root". This option applies to job allocations.

If user namespaces are active, then the mapping of users in the namespace must match the same namespace as MUNGE. If it doesn't, the job will be rejected by slurmd. 💀

• -k, --no-kill[=off] — don't automatically terminate a job if one of the nodes it's been allocated fails. This option applies to job and step allocations. The job assumes all responsibilities for fault-tolerance. Tasks launched with this option will not be considered terminated (e.g. -K, --kill-on-bad-exit and -W, --wait options will have no effect upon the job step). The active job step (MPI job) will likely suffer a fatal error, but subsequent job steps may be run if this option is specified. 🫶

Specify an optional argument of "off" to disable the effect of the SLURM_NO_KILL environment variable.

The default action is to terminate the job upon node failure.

• -F, --nodefile=<node_file> — much like --nodelist, but the list lives inside a file named node file. The node names in the list may also span multiple lines in the file. Duplicate node names in the file will be ignored. The order of the node names in the list is not important; the node names will be sorted by Slurm.
• -w, --nodelist={<node_name_list>|<filename>} — request a specific list of nodes. The job will contain as many of these nodes as possible based on the resource requirements, delaying execution as needed to wait for resources to become available.

If you specify a minimum node or processor count larger than what the supplied node list can satisfy, additional resources will be allocated on other nodes as needed (unless --segment is also used). Conversely, a lower node or processor count may only require a subset of the supplied node list.

The list may be specified as a comma-separated list of nodes, a range of nodes (e.g. node[1-5,7,...]), or a filename. The nodes list will be assumed to be a filename if it contains a "/" character. The order of the node names in the list is not important; the node names will be sorted by Slurm.

Repeated node names may be useful when using --distribution=arbitrary. An asterisk and a repetition count may be appended to a node name. For example, "node1,node1" and "node1*2" are equivalent. This option applies to job and step allocations. 🌈

• -N, --nodes=<minnodes>[-maxnodes]|<size_string> — request that a minimum of minnodes nodes be allocated to this job. A maximum node count may also be specified with maxnodes. If only one number is specified, it's used as both the minimum and the maximum node count. Node count can also be given as a size_string. The size_string specification identifies what nodes values should be used. Multiple values may be specified using a comma separated list or with a step function by a suffix containing a colon and number values with a "-" separator. For example, "--nodes=1-15:4" is equivalent to "--nodes=1,5,9,13". The partition's node limits supersede those of the job. If a job's node limits are outside of the range permitted for its associated partition, the job will be left in a PENDING state. That's on purpose — it permits possible execution at a later time, when the partition limit is changed. If a job node limit exceeds the number of nodes configured in the partition, the job will be rejected. Note that the environment variable SLURM_JOB_NUM_NODES (and SLURM_NNODES for backwards compatibility) will be set to the count of nodes actually allocated to the job. See the ENVIRONMENT VARIABLES section for more info. If -N is not specified, the default behavior is to allocate enough nodes to satisfy the requested resources as expressed by per-job specification options, e.g. -n, -c, --ntasks-per-node and --gpus. ✨

This derivation applies to job allocations only. When -N is omitted, the derived count is --ntasks, or --ntasks divided by --ntasks-per-node and rounded up if that option is given. Without --ntasks-per-node the count is only an upper bound, so MaxNodes reduces it rather than rejecting the job. With --ntasks-per-node the count is the number of nodes the job requires, and MaxNodes rejects it as it would an explicit count. Either way, a job whose tasks do not fit is rejected or left in a PENDING state.

The derived count is not lower than the partition's MinNodes at submit time. Raising the partition's MinNodes later does not re-derive it, so a job whose count is below the new value stays PENDING until its node count is set explicitly. 😭

The job will be allocated as many nodes as possible within the range specified and without delaying the initiation of the job. If the number of tasks is given and a number of requested nodes is also given, the number of nodes used from that request will be reduced to match that of the number of tasks if the number of nodes in the request is greater than the number of tasks. The node count specification may include a numeric value followed by a suffix of "k" (multiplies numeric value by 1,024) or "m" (multiplies numeric value by 1,048,576). This option applies to job and step allocations.

NOTE: This option cannot be used with arbitrary distribution. 🚫

• -n, --ntasks=<number> — specify the number of tasks to run. Request that srun allocate resources for ntasks tasks. For a job allocation, if the --nodes option is not used, this option is used to implicitly determine the node count. The default is one task per node, but note that the --cpus-per-task option will change this default. This option applies to job and step allocations. 💖
• --ntasks-per-core=<ntasks> — request the maximum ntasks be invoked on each core. This option applies to job and step allocations. Meant to be used with the --ntasks option. Related to --ntasks-per-node except at the core level instead of the node level. If set to 1, it will imply --cpu-bind=cores. Otherwise, if set to a value greater than 1, it will imply --cpu-bind=threads. Automatic binding behavior can be avoided by also specifying --cpu-bind=none. Slurm may allocate more cpus than what was requested in order to respect this option. 🧚

NOTE: This option is not supported when using SelectType=select/linear. This value can not be greater than --threads-per-core.

• --ntasks-per-gpu=<ntasks> — bestie, this one's a whole mood: it asks for ntasks tasks for every GPU. It works two ways. ✨ Way 1: you also pass --ntasks, and then a type-less GPU spec gets auto-figured-out to satisfy --ntasks-per-gpu. Way 2: you specify the GPUs you want (like via --gpus or --gres) but don't pass --ntasks, and the total task count gets automatically determined for you. The CPU count will automatically be bumped up if needed to allow for whatever calculated task count comes out. Heads up: this option implicitly sets --tres-bind=gres/gpu:single:<ntasks> — but you can override that with an explicit --tres-bind=gres/gpu spec. 🫶 Not compatible with a node range (that's the -N<minnodes-maxnodes> situation). Also not compatible with --gpus-per-task, --gpus-per-socket, or --ntasks-per-node. And it's only supported if SelectType=cons_tres is configured — either directly or indirectly on Cray systems. 💅
• --ntasks-per-node=<ntasks> — requests that ntasks get invoked on each node. 🌸 If you use it together with --ntasks, then --ntasks wins and --ntasks-per-node gets treated as a maximum count of tasks per node. It's meant to go with the --nodes option. If --nodes isn't used but --ntasks is given, this option determines the job's node count. See -N, --nodes. This is related to --cpus-per-task=ncpus, but you don't need to know the actual number of cpus on each node — convenient, right? Sometimes you just want to say "no more than this many tasks per node," and honestly iconic use cases include: submitting a hybrid MPI/OpenMP app where only one MPI "task/rank" should land on each node while the OpenMP part uses all the parallelism the node has, or submitting a single setup/cleanup/monitoring job to each node of a pre-existing allocation as one step in a bigger job script. This option applies to job allocations. 🎀
• --ntasks-per-socket=<ntasks> — asks for a maximum of ntasks invoked on each socket. 🦋 This one applies to the job allocation, but not to step allocations. Meant to be used with the --ntasks option. It's related to --ntasks-per-node except at socket level instead of node level. Masks will automatically be generated to bind the tasks to specific sockets — unless you specify --cpu-bind=none.

NOTE: this option is not supported when using SelectType=select/linear. 💀

• --oom-kill-step[={0|1}] — decides whether to kill the entire step if an OOM event is detected in any task of the step. 😭 It overwrites the "OOMKillStep" setting in TaskPluginParam from slurm.conf and the allocation settings. When unset, it uses the setting in slurm.conf. When set, a value of "0" disables killing the entire step, and "1" enables it. Default is "1" (enabled) when the option shows up with no value. 🔥
• --open-mode={append|truncate} — opens the output and error files in append or truncate mode, as specified. ✨ For heterogeneous job steps the default value is "append". Otherwise the default is whatever the system configuration parameter JobFileAppend says. This option applies to job and step allocations. See EXAMPLE below. 👀
• -o, --output=<filename_pattern> — sets the "filename pattern" for stdout redirection. 🎀 By default, in interactive mode, srun collects stdout from all tasks and sends it over TCP/IP to your attached terminal. With --output, stdout can go to a file, to one file per task, or to /dev/null. Check the IO Redirection section below for all the various forms of filename pattern. And ngl — if the specified file already exists, it gets overwritten. 😳 If --error isn't also on the command line, then both stdout and stderr get directed to the file from --output. This option applies to job and step allocations. 💖
• -O, --overcommit — overcommit resources, babes. 🫶 This option applies to job and step allocations. When it's applied to a job allocation (not counting jobs requesting exclusive access to the nodes), resources get allocated as if only one task per node was requested. That means the requested cpus per task (-c, --cpus-per-task) are allocated per node rather than being multiplied by the number of tasks. Options you'd use to specify the number of tasks per node, socket, core, etc. are just ignored. 💅 When it's applied to job step allocations (i.e. the srun command run inside an existing job allocation), this option lets you launch more than one task per CPU. Normally srun won't allocate more than one process per CPU — by specifying --overcommit you're explicitly allowing more than one process per CPU. But no more than MAX_TASKS_PER_NODE tasks are permitted to execute per node.

NOTE: MAX_TASKS_PER_NODE is defined in the file slurm.h and it is not a variable — it's set at Slurm build time. 🧚

• --overlap — specifying --overlap lets steps share all resources (CPUs, memory, and GRES) with all other steps. 🌈 A step using this option will overlap all other steps — even the ones that didn't specify --overlap. By default, steps do not share resources with other parallel steps. This option applies to step allocations. ⭐
• -s, --oversubscribe — the job allocation can over-subscribe resources with other running jobs. 🔥 The resources being over-subscribed can be nodes, sockets, cores, and/or hyperthreads, depending on configuration. The default over-subscribe behavior depends on system configuration, and the partition's OverSubscribe option takes precedence over the job's option. This option may get your allocation granted sooner than if --oversubscribe wasn't set, and it allows higher system utilization — but application performance will likely suffer from all that competition for resources. 😭 This option applies to job allocations.

NOTE: this option is mutually exclusive with --exclusive. Periodt.

• --parsable — outputs only the step ID (<jobid>.<stepid>) on a single line of stdout. ✨ It suppresses the leading "Submitted step " text. Analogous to sbatch --parsable, but it includes a step ID alongside the job ID. Errors are still printed normally to stderr. Requires --async. And note that --quiet suppresses the step ID along with the other informational messages. 👑
• -p, --partition=<partition_names> — requests a specific partition for the resource allocation. 🎀 If you don't specify one, the default behavior is to let the slurm controller pick the default partition as designated by the system administrator. If the job can use more than one partition, list their names in a comma-separated list — and the one offering earliest initiation gets used, with no regard given to the partition name ordering (although higher priority partitions are considered first). When the job is initiated, the name of the partition used is placed first in the job record partition string. This option applies to job allocations. 💖
• --prefer=<list> — nodes can have features assigned to them by the Slurm administrator. 🦋 With the prefer option, users can say which of those features are desired but not required by their job. This option operates independently from --constraint and will override whatever is set there if possible. When scheduling, the features in --prefer get tried first — and if a node set isn't available with those features, then --constraint is attempted. See --constraint for more info; this option behaves the same way. ✨
• -E, --preserve-env — passes the current values of the environment variables SLURM_JOB_NUM_NODES and SLURM_NTASKS through to the executable, rather than computing them from command line parameters. 💅 This option applies to job allocations.
• --priority=<value> — requests a specific job priority. 🎀 May be subject to configuration-specific constraints. The value should either be a numeric value or "TOP" (for the highest possible value). Only Slurm operators and administrators can set the priority of a job. This option applies to job allocations only. 👑
• --profile={all|none|<type>[,<type>...]} — enables detailed data collection by the acct_gather_profile plugin. ✨ Detailed data are typically time-series stored in an HDF5 file for the job, or in an InfluxDB database, depending on the configured plugin. This option applies to job and step allocations.
  • All — all data types are collected. (Cannot be combined with other values.)
  • None — no data types are collected. This is the default. (Cannot be combined with other values.)

Valid type values are:

• Energy — Energy data is collected. ⚡
• Task — Task (I/O, Memory, ...) data is collected. 🧚
• Filesystem — Filesystem data is collected. 🍓
• Network — Network (InfiniBand) data is collected. 🌈

• --prolog=<executable> — srun will run executable just before launching the job step. 💖 The command line arguments for executable will be the command and arguments of the job step. If executable is "none", then no srun prolog will be run. This parameter overrides the SrunProlog parameter in slurm.conf. And it's completely independent from the Prolog parameter in slurm.conf. This option applies to job allocations. 🎀
• --propagate[=rlimit[,rlimit...]] — lets users specify which of the modifiable (soft) resource limits get propagated to the compute nodes and applied to their jobs. 🦋 If no rlimit is specified, then all resource limits get propagated. The following rlimit names are supported by Slurm (although some options may not be supported on some systems):
  • ALL — all limits listed below (default) ⭐
  • NONE — no limits listed below 💀
  • AS — the maximum address space (virtual memory) for a process. 🧠
• CORE — the max size of a core file 💾
  • CPU — the max amount of CPU time ⏳
  • DATA — the max size of a process's data segment
  • FSIZE — the max size of files created. Heads up bestie: if the user sets FSIZE to less than the current size of the slurmd.log, job launches will fail with a 'File size limit exceeded' error. 😭 not the log file eating your job
  • MEMLOCK — the max size that may be locked into memory
  • NOFILE — the max number of open files
  • NPROC — the max number of processes available
  • RSS — the max resident set size. Ngl this one only has effect with Linux kernels 2.4.30 or older, or BSD. vintage behaviour fr
  • STACK — the max stack size

This option applies to job allocations.

• --pty, --pty=<File Descriptor> — run task zero in pseudo terminal mode, or using the pseudo terminal that <File Descriptor> specifies. It implicitly sets --unbuffered, and it implicitly sets --error and --output to /dev/null for every task except task zero — which lowkey may cause those other tasks to exit immediately (like, shells will typically exit immediately in that situation, which is giving 💀). This option applies to step allocations.
• -q, --qos=<qos> — request a quality of service for the job, or a comma separated list of QOS. If you're requesting a list, it gets ordered by the priority of the QOS you gave, first one = highest priority. 👑 QOS values can be defined for each user/cluster/account association in the Slurm database. Users get limited to their association's defined set of qos's when the Slurm configuration parameter AccountingStorageEnforce includes "qos" in its definition. This option applies to job allocations.
• -Q, --quiet — suppress informational messages from srun. Errors will still be displayed, so don't think you can hide from consequences bestie 🙈. This option applies to job and step allocations.
• --quit-on-interrupt — quit immediately on a single SIGINT (Ctrl+C). Using this disables the status feature you'd normally get when srun receives a single Ctrl+C, and instead makes srun immediately terminate the running job. abrupt much? ✨ This option applies to step allocations.
• --reboot — force the allocated nodes to reboot before starting the job. This is only supported with some system configurations and is otherwise silently ignored. Only root, SlurmUser or admins can reboot nodes. This option applies to job allocations.
• -r, --relative=<n> — run a job step relative to node n of the current allocation. Perfect for spreading several job steps out among the nodes of the current job 🦋. If -r is used, the current job step begins at node n of the allocated nodelist, where the first node is node 0. The -r option is not permitted with -w or -x, and it will result in a fatal error when you're not running within a prior allocation (i.e. when SLURM_JOB_ID is not set). The default for n is 0. If the value of --nodes exceeds the number of nodes identified with the --relative option, a warning message gets printed and --relative takes precedence. This option applies to step allocations.
• --reservation=<reservation_names> — allocate resources for the job from the named reservation. If the job can use more than one reservation, specify their names in a comma separate list and the one offering earliest initiation wins 🎀. Each reservation is considered in the order it was requested. All reservations will be listed in scontrol/squeue through the life of the job. In accounting the first reservation will be seen, and after the job starts the reservation actually used will replace it.
• --resources=<resource_names> — specification of hierarchical resources which must be allocated to this job. Resource names can be followed by a colon and count (the default count is one). Multiple resources in Mode 1 and Mode 2 can be requested in a comma separated list, but only one of Mode 3 can be requested. For example, "--resources=flat:2,natural:1".

See https://slurm.schedmd.com/hres.html for more information on use of Hierarchical Resource with Slurm. iykyk 💅

• --resv-ports[=count] — reserve communication ports for this job. Users can specify the number of ports they want to reserve. The parameter MpiParams=ports=12000-12999 must be specified in slurm.conf. If the number of reserved ports is zero then no ports are reserved. Used for native Cray's PMI only. This option applies to job and step allocations.
• --segment=<segment_size> — when a block or ring topology is used, this defines the size of the segments that will be used to create the job allocation. No requirement would be placed on all segments for a job needing to be placed within the same higher-level block.

NOTE: if the requested node count (--nodes) is larger than the requested segment size, it must also be evenly divisible by the segment size. If all nodes fit within a single segment, this option has no effect. 🧚

NOTE: when used in conjunction with --nodelist=<node_list>, the requested node count must be less than or equal to the total number of unique nodes specified in the --nodelist argument. Requesting more nodes than available unique nodes in the provided --nodelist will result in the job being rejected by slurmctld. rejected, blocked, ghosted 💀

• --send-libs[=yes|no] — if set to yes (or given no argument), autodetect and broadcast the executable's shared object dependencies to the allocated compute nodes. The files are placed in a directory alongside the executable. LD_LIBRARY_PATH is automatically updated to include this cache directory as well. This overrides the default behavior configured in slurm.conf SbcastParameters send_libs. This option only works in conjunction with --bcast. See also --bcast-exclude. 🫶
• --signal=[R:]<sig_num>[@sig_time] — when a job is within sig_time seconds of its end time, send it the signal sig_num. Due to the resolution of event handling by Slurm, the signal may be sent up to 60 seconds earlier than specified 😭. sig_num may either be a signal number or name (e.g. "10" or "USR1"). sig_time must have an integer value between 0 and 65535. By default, no signal is sent before the job's end time. If a sig_num is specified without any sig_time, the default time will be 60 seconds. This option applies to job allocations.

Use the "R:" option to allow this job to overlap with a reservation with MaxStartDelay set. If the "R:" option is used, preemption must be enabled on the system, and if the job is preempted it will be requeued if allowed, otherwise the job will be canceled. To have the signal sent at preemption time see the send_user_signal PreemptParameter.

• --slurmd-debug=<level> — specify a debug level for this step. The level may be specified either as an integer value between 2 [error] and 6 [debug2], or as one of the SlurmdDebug tags. 🔍
  • error — log only errors
  • info — log errors and general informational messages
  • verbose — log errors and verbose informational messages
  • debug — log errors and verbose informational messages and debugging messages
  • debug2 — log errors and verbose informational messages and more debugging messages

The slurmd debug information is copied onto the stderr of the job. By default only errors are displayed. This option applies to job and step allocations.

• --sockets-per-node=<sockets> — restrict node selection to nodes with at least the specified number of sockets. See additional information under the -B option above when the task/affinity plugin is enabled. This option applies to job allocations.

NOTE: this option may implicitly impact the number of tasks if -n was not specified. sneaky little side effect 👀

• --spread-job — spread the job allocation over as many nodes as possible and attempt to evenly distribute tasks across the allocated nodes. This option disables the topology/tree plugin. This option applies to job allocations. go touch some nodes 🌈
• --spread-segments — prevent nodes within the same base block from being allocated to separate segments within the same block.

This option applies to job allocations.

NOTE: this option will only work with the topology/block plugin.

• --stepmgr — enable slurmstepd step management per-job if it isn't enabled system wide. This enables job steps to be managed by a single extern slurmstepd associated with the job to manage steps. This is beneficial for jobs that submit many steps inside their allocations 🎀. PrologFlags=contain must be set. This option applies to job allocations.
• --switches=<count>[@max-time] — when a tree topology is used, this defines the maximum count of leaf switches desired for the job allocation, and optionally the maximum time to wait for that number of switches. If Slurm finds an allocation containing more switches than the count specified, the job remains pending until it either finds an allocation with the desired switch count or the time limit expires. If there is no switch count limit, there is no delay in starting the job. ⭐

Acceptable time formats include "minutes", "minutes:seconds", "hours:minutes:seconds", "days-hours", "days-hours:minutes" and "days-hours:minutes:seconds". The job's maximum time delay may be limited by the system administrator using the SchedulerParameters configuration parameter with the max_switch_wait parameter option. On a dragonfly network the only switch count supported is 1, since communication performance will be highest when a job is allocated resources on one leaf switch or more than 2 leaf switches. 🐉 The default max-time is the max_switch_wait SchedulerParameters. This option applies to job allocations.

• --task-epilog=<executable> — bestie, the slurmstepd daemon runs executable right after each task wraps up, and the ordering tea is important: this fires before any TaskEpilog parameter from slurm.conf gets its turn ✨. Keep it short-lived, okay? If it doesn't terminate within a few seconds it gets killed, and any descendant processes go down with it 😭. This one applies to step allocations.
• --task-prolog=<executable> — the slurmstepd daemon runs executable just before launching each task, and this one lands after any TaskProlog parameter from slurm.conf has run 🫶. Besides the usual environment variables, it gets SLURM_TASK_PID so it can identify the process ID of the task being started 👑. And the cute part: standard output from this program in the form export NAME=value gets used to set environment variables for the task being spawned 🌸. Applies to step allocations too.
• --test-only — returns an estimate of when a job would be scheduled to run, given the current job queue and all the other srun arguments specifying the job 🔮. It limits srun's behavior to just returning information — no job is actually submitted, no cap. Applies to job allocations. You can also combine it with --jobid to get the estimated start time of a pending job ⭐.
• --thread-spec=<num> — the count of specialized threads per node reserved by the job for system operations, not used by the application. The app won't use these threads but will be charged for their allocation 💸. Can't be used with the --core-spec option. Applies to job allocations.

NOTE: explicitly setting a job's specialized thread value implicitly sets its --exclusive option, reserving entire nodes for the job. Like, the whole node. Yours. Iconic but expensive 🎀

• -T, --threads=<nthreads> — limits the number of concurrent threads used to send the job request from the srun process to the slurmd processes on the allocated nodes 🧵. Default is one thread per allocated node, up to a max of 60 concurrent threads. Specifying this limits the concurrent threads to nthreads (less than or equal to 60). Honestly this should only be used to set a low thread count for testing on very small memory computers — it's giving "tiny machine, big dreams" vibes 💀.
• --threads-per-core=<threads> — restricts node selection to nodes with at least the specified number of threads per core 🌈. In task layout, uses the specified maximum number of threads per core. Implies --cpu-bind=threads unless overridden by command line or environment options.

NOTE: "Threads" here refers to the number of processing units on each core, not the number of application tasks to be launched per core. See the additional info under the -B option above when the task/affinity plugin is enabled. This option applies to job and step allocations. NOTE: this option may implicitly impact the number of tasks if -n was not specified — so like, watch out bestie 🙈

• -t, --time=<time> — sets a limit on the total run time of the job allocation ⏰. If the requested time limit exceeds the partition's time limit, the job gets left in a PENDING state (possibly indefinitely, and yes that's as tragic as it sounds 😭). The default time limit is the partition's default time limit. When the time limit is reached, each task in each job step is sent SIGTERM followed by SIGKILL. The interval between those signals is specified by the Slurm configuration parameter KillWait. The OverTimeLimit configuration parameter may permit the job to run longer than scheduled. Time resolution is one minute, and second values are rounded up to the next minute 🕐.

A time limit of zero requests that no time limit be imposed — free bird energy 🦋. Acceptable time formats include "minutes", "minutes:seconds", "hours:minutes:seconds", "days-hours", "days-hours:minutes" and "days-hours:minutes:seconds". Applies to job and step allocations.

• --time-min=<time> — sets a minimum time limit on the job allocation ✨. If specified, the job may have its --time limit lowered to a value no lower than --time-min, if doing so permits the job to begin execution earlier than otherwise possible. This is done by a backfill scheduling algorithm to allocate resources otherwise reserved for higher priority jobs or an upcoming reservation 👑. The lowered limit is set when the job is allocated resources and is not changed afterwards, periodt. Acceptable time formats include "minutes", "minutes:seconds", "hours:minutes:seconds", "days-hours", "days-hours:minutes" and "days-hours:minutes:seconds". Applies to job allocations.
• --tmp=<size>[units] — specifies a minimum amount of temporary disk space per node 💾. Default units are mebibytes. Different units can be specified using the suffix [K|M|G|T]. Applies to job allocations.
• --treewidth=<size> — specifies the width of the fanout 🌳. The default is the TreeWidth specified in slurm.conf. The value may not exceed 65533. A value of "off" disables the fanout entirely — bye fanout 👋.
• --tres-bind=<tres>:[verbose,]<type>[+<tres>:[verbose,]<type>...] — specifies a list of tres with their task binding options 🎀. Currently gres are the only supported tres for this option. Specify gres as "gres/<gres_name>" (e.g. gres/gpu).

Example:

--tres-bind=gres/gpu:verbose,map:0,1,2,3+gres/nic:closest

By default, most tres are not bound to individual tasks.

Supported binding type options for gres:

• closest — bind each task to the gres(s) which are closest 📍. In a NUMA environment, each task may be bound to more than one gres (i.e. all gres in that NUMA environment).
• map:<list> — bind by setting gres masks on tasks (or ranks) as specified, where <list> is <gres_id_for_task_0>,<gres_id_for_task_1>,... 🗺️. gres IDs are interpreted as decimal values. If the number of tasks (or ranks) exceeds the number of elements in this list, elements in the list will be reused as needed starting from the beginning of the list. To simplify support for large task counts, the lists may follow a map with an asterisk and repetition count. For example "map:0*4,1*4". If the task/cgroup plugin is used and ConstrainDevices is set in cgroup.conf, then the gres IDs are zero-based indexes relative to the gres allocated to the job (e.g. the first gres is 0, even if the global ID is 3). Otherwise, the gres IDs are global IDs, and all gres on each node in the job should be allocated for predictable binding results.
• mask:<list> — bind by setting gres masks on tasks (or ranks) as specified, where <list> is <gres_mask_for_task_0>,<gres_mask_for_task_1>,... 🎭. The mapping is specified for a node and identical mapping is applied to the tasks on every node (i.e. the lowest task ID on each node is mapped to the first mask specified in the list, etc.). gres masks are always interpreted as hexadecimal values but can be preceded with an optional 0x. To simplify support for large task counts, the lists may follow a map with an asterisk and repetition count. For example "mask:0x0f*4,0xf0*4". If the task/cgroup plugin is used and ConstrainDevices is set in cgroup.conf, then the gres IDs are zero-based indexes relative to the gres allocated to the job (e.g. the first gres is 0, even if the global ID is 3). Otherwise, the gres IDs are global IDs, and all gres on each node in the job should be allocated for predictable binding results.
• none — do not bind tasks to this gres (turns off implicit binding from --tres-per-task and --gpus-per-task) 🚫.
• per_task:<gres_per_task> — each task will be bound to the number of gres specified in <gres_per_task> 💅. Tasks are preferentially assigned gres with affinity to cores in their allocation, like in closest, though they will take any gres if they are unavailable. If no affinity exists, the first task will be assigned the first x number of gres on the node etc. Shared gres will prefer to bind one sharing device per task if possible.
• single:<tasks_per_gres> — like closest, except that each task can only be bound to a single gres, even when it can be bound to multiple gres that are equally close 🎯. The gres to bind to is determined by <tasks_per_gres>, where the first <tasks_per_gres> tasks are bound to the first gres available, the second <tasks_per_gres> tasks are bound to the second gres available, etc. This is basically a block distribution of tasks onto available gres, where the available gres are determined by the socket affinity of the task and the socket affinity of the gres as specified in gres.conf's Cores parameter.

NOTE: shared gres binding is currently limited to per_task or none 🫶

• --tres-per-task=<list> — bestie, this one lets you name exactly which trackable resources each task needs inside the job's resource allocation. You give it a comma-delimited list, and every entry looks like "trestype[/tresname]=count". The trestype is the kind of trackable resource you're asking for (cpu, gres, license, etc.). The tresname is the name of that resource — you can peek at the names with sacctmgr show tres. It's required whenever a name exists for that tres type, like gres, license, etc. (think gpu, gpu:a100). Wanna request a license with this? Then the license(s) must be defined in the AccountingStorageTRES parameter of slurm.conf. The count is literally how many of those resources.

And the count can carry a suffix, periodt:

• "k" or "K" — multiple of 1024
• "m" or "M" — multiple of 1024 x 1024
• "g" or "G" — multiple of 1024 x 1024 x 1024
• "t" or "T" — multiple of 1024 x 1024 x 1024 x 1024
• "p" or "P" — multiple of 1024 x 1024 x 1024 x 1024 x 1024

Examples, because we love a visual 🎀:

--tres-per-task=cpu=4
--tres-per-task=cpu=8,license/ansys=1
--tres-per-task=gres/gpu=1
--tres-per-task=gres/gpu:a100=2

The specified resources get allocated to the job on each node. Which trackable resources even exist is up to your system administrator.

NOTE: using this option with gres/gpu or gres/shard will implicitly set --tres-bind=gres/[gpu|shard]:per_task:<tres_per_task> — or, if multiple gpu types are specified, --tres-bind=gres/gpu:per_task:<gpus_per_task_type_sum>. You can override that with an explicit --tres-bind spec. No cap.

NOTE: invalid TRES for --tres-per-task include bb,billing,energy,fs,mem,node,pages,vmem. Don't try it 💀

• -u, --unbuffered — okay so by default, the connection between slurmstepd and your user-launched application runs over a pipe. The stdio output the app writes gets buffered by glibc until it's flushed or the output is set as unbuffered — see setbuf(3) if you're curious. Specify this option and the tasks execute with a pseudo terminal instead, so the app output is unbuffered. This one applies to step allocations ✨
• --usage — display a brief help message and exit. Lil snack-sized help 🍓
• --use-min-nodes — if a range of node counts is given, prefer the smaller count. Lowkey the chill option.
• -v, --verbose — turn up the volume on srun's informational messages 🔊 Multiple -v's will increase the verbosity even further. By default only errors get displayed. This option applies to job and step allocations.
• -V, --version — display version information and exit. Iconic.
• --wait-for-children — wait for all the processes in each task to finish before considering that task ended. Without this option, the default behavior is to only wait for the parent process in each task to finish. The way that this interacts with --kill-on-bad-exit matters, so listen up:
  • If --kill-on-bad-exit=1 and the parent process exits with a non-zero exit code → the task will end.
  • If --kill-on-bad-exit=0 and the parent process exits with a non-zero exit code → the task continues until all children processes have exited.

This option requires proctrack/cgroup and cgroup/v2. No substitutions, bestie 🙈

• -W, --wait=<seconds> — how long do we wait after the first task terminates before terminating all the remaining tasks? That's the question this answers. A value of 0 means an unlimited wait — and heads up, a warning will be issued after 60 seconds. The default value is set by the WaitTime parameter in the slurm configuration file (see slurm.conf(5)). This is genuinely useful for making sure a job gets terminated in a timely fashion if one or more tasks terminate prematurely.

Note: -K, --kill-on-bad-exit takes precedence over -W, --wait — it'll terminate the job immediately if a task exits with a non-zero exit code. This option applies to job allocations.

• --wckey=<wckey> — specify the wckey to be used with the job. If TrackWCKey=no (the default) in the slurm.conf, this value is just ignored. This option applies to job allocations.
• --x11[={all|first|last}] — sets up X11 forwarding on "all", "first" or "last" node(s) of the allocation. This option is only enabled if Slurm was compiled with X11 support and PrologFlags=x11 is defined in the slurm.conf. Default is "all" 🖥️✨

NOTE: this option may not work if --clusters is also provided. Such jobs will be accepted (with a warning), but X11 forwarding is not guaranteed to work. F around and find out, respectfully 💀

---

srun will submit the job request to the slurm job controller, then initiate all processes on the remote nodes. If the request can't be met immediately, srun will block until the resources are free to run the job — yes, it will just sit there, patient main character energy. If the -I (--immediate) option is specified, srun will terminate instead when resources aren't immediately available.

When initiating remote processes, srun will propagate the current working directory — unless --chdir=<path> is specified, in which case path becomes the working directory for the remote processes. Simple as that 🫶

The -n, -c, and -N options control how CPUs and nodes get allocated to the job. Specify only the number of processes to run with -n and you get a default of one CPU per process. Specify the number of CPUs required per task (-c) and you can allocate more than one CPU per process. Specify the number of nodes with -N and srun will attempt to allocate at least that many nodes.

Combinations of the above three options can change how processes are distributed across nodes and cpus. For instance, specifying both the number of processes and the number of nodes to run on implies the number of processes per node. But if the number of CPUs per process matters more to you, then specify the number of processes (-n) and the number of CPUs per process (-c). Slay, math 🧮

srun will refuse to allocate more than one process per CPU unless --overcommit (-O) is also specified. Firm boundary. Love that for it.

srun will attempt to meet the above specifications "at a minimum." That is: if 16 nodes are requested for 32 processes, and some nodes don't have 2 CPUs, the allocation of nodes will be increased in order to meet the demand for CPUs. In other words, a minimum of 16 nodes is being requested. However — and this is the plot twist — if 16 nodes are requested for 15 processes, srun will consider this an error, because 15 processes cannot run across 16 nodes. Math said no 😭

🎀 IO Redirection 🎀

By default, stdout and stderr get redirected from all tasks to the stdout and stderr of srun, and stdin gets redirected from the standard input of srun to all remote tasks. If stdin is only to be read by a subset of the spawned tasks, specifying a file to read from rather than forwarding stdin from the srun command may be preferable — because it avoids moving and storing data that will never be read. Efficiency queen 👑

For OS X, the poll() function does not support stdin, so input from a terminal is not possible. Pour one out 🫗

This behavior may be changed with the --output, --error, and --input (-o, -e, -i) options. Note that --error won't redirect the stderr of srun itself — only the stderr from the tasks. Valid format specifications for these options are:

• all — stdout stderr is redirected from all tasks to srun. stdin is broadcast to all remote tasks. (This is the default behavior.)
• none — stdout and stderr is not received from any task. stdin is not sent to any task (stdin is closed).
• taskid — stdout and/or stderr are redirected from only the task with relative id equal to taskid, where 0 <= taskid <= ntasks, where ntasks is the total number of tasks in the current job step. stdin is redirected from the stdin of srun to this same task. This file will be written on the node executing the task.
• filename — srun will redirect stdout and/or stderr to the named file from all tasks. stdin will be redirected from the named file and broadcast to all tasks in the job. filename refers to a path on the host that runs srun. Depending on the cluster's file system layout, this may result in the output appearing in different places depending on whether the job is run in batch mode. Awkward but true 🙈
• filename pattern — srun allows a filename pattern to be used to generate the named IO file described above. The following list of format specifiers may be used in the format string to generate a filename that will be unique to a given jobid, stepid, node, or task. In each case, the appropriate number of files are opened and associated with the corresponding tasks. Note that any format string containing %t, %n, and/or %N will be written on the node executing the task rather than the node where srun executes.
  • \\ — Do not process any of the replacement symbols.
  • %% — The character "%".
  • %A — Job array's master job allocation number.
  • %a — Job array ID (index) number.
  • %J — jobid.stepid of the running job (e.g. "128.0"). The stepid is only expanded for regular steps, not for special steps like "batch" or "extern".
  • %j — jobid of the running job.
  • %S — SLUID of the running job.
  • %s — stepid of the running job.
  • %N — short hostname. This will create a separate IO file per node.
  • %n — Node identifier relative to current job (e.g. "0" is the first node of the running job). This will create a separate IO file per node.
  • %t — task identifier (rank) relative to current job. This will create a separate IO file per task.
  • %u — User name.
  • %x — Job name.

A number placed between the percent character and the format specifier may be used to zero-pad the result in the IO filename to a minimum of the specified numbers. This number is ignored if the format specifier corresponds to non-numeric data (%N, for example). The maximal number is 10 — if a value greater than 10 is used, the result is padding up to 10 characters. Some examples of how the format string may be used for a 4 task job step with a JobID of 128 and step id of 0 are included below:

job%J.out      job128.0.out

                 job%4j.out     job0128.out

                 job%2j-%2t.out job128-00.out, job128-01.out, ...

🔥 PERFORMANCE 🔥

Okay so real talk for a sec, bestie — this one's important and it's giving cautionary tale.

Every time you run srun, it sends a remote procedure call over to slurmctld. And like... if enough calls from srun (or any other Slurm client command that sends RPCs to the slurmctld daemon) all show up at once? Yeah, that can degrade the performance of the slurmctld daemon, and it could even straight-up result in a denial of service. 😭 Not cute. Not the vibe.

So please, do NOT run srun or other Slurm client commands that send RPCs to slurmctld from loops in shell scripts or other programs. Fr fr. If your program is calling srun, keep those calls limited to the bare minimum necessary for the information you're trying to gather. Your cluster will thank you. 🫶

---

💖 INPUT ENVIRONMENT VARIABLES 💖

When it starts up, srun reads and handles the options set in the environment variables listed below. Most of these bad boys get set the same way the options are set, as defined up above. ✨

So here's the tea for flag options that are defined to expect no argument: you can enable the option by setting the environment variable without a value (empty or NULL string), or to the string 'yes', or to a non-zero number. Any other value for that environment variable means the option just... doesn't get set. 🙈 There are a couple exceptions to these rules, and those get noted below, iykyk.

NOTE: Command line options always override environment variable settings. Always always. Periodt. 👑

• PMI_FANOUT — Used exclusively with PMI (MPICH2 and MVAPICH2), this controls the fanout of data communications. The srun command sends messages to application programs (via the PMI library), and those applications may be called upon to forward that data to up to this number of additional tasks. Higher values offload work from the srun command to the applications, and likely increase the vulnerability to failures. Default value is 32. 🍓
• PMI_FANOUT_OFF_HOST — Also used exclusively with PMI (MPICH2 and MVAPICH2), controls the fanout of data communications. The srun command sends messages to application programs (via the PMI library) and those applications may be called upon to forward that data to additional tasks. By default, srun sends one message per host, and one task on that host forwards the data to other tasks on that host up to PMI_FANOUT. If PMI_FANOUT_OFF_HOST is defined, the user task may be required to forward the data to tasks on other hosts. Setting PMI_FANOUT_OFF_HOST may increase performance — but since more work is performed by the PMI library loaded by the user application, failures also can be more common and more difficult to diagnose. 😬 Should be disabled/enabled by setting to 0 or 1.
• PMI_TIME — Used exclusively with PMI (MPICH2 and MVAPICH2), this controls how much the communications from the tasks to the srun are spread out in time, so we don't overwhelm the srun command with work. Default value is 500 (microseconds) per task. On relatively slow processors, or systems with very large processor counts (and large PMI data sets), higher values may be required. ✨
• SLURM_ACCOUNT — Same as -A, --account
• SLURM_ACCTG_FREQ — Same as --acctg-freq
• SLURM_BCAST — Same as --bcast
• SLURM_BCAST_EXCLUDE — Same as --bcast-exclude
• SLURM_BURST_BUFFER — Same as --bb
• SLURM_CLUSTERS — Same as -M, --clusters
• SLURM_COMPRESS — Same as --compress
• SLURM_CONF — The location of the Slurm configuration file.
• SLURM_CONSTRAINT — Same as -C, --constraint
• SLURM_CORE_SPEC — Same as --core-spec
• SLURM_CPU_BIND — Same as --cpu-bind
• SLURM_CPU_FREQ_REQ — Same as --cpu-freq.
• SLURM_CPUS_PER_GPU — Same as --cpus-per-gpu
• SLURM_CPUS_PER_TASK — Same as -c, --cpus-per-task or --tres-per-task=cpu=#
• SLURM_DEBUG — Same as -v, --verbose, when set to 1, when set to 2 gives -vv, etc.
• SLURM_DEBUG_FLAGS — Specify debug flags for srun to use. See DebugFlags in the slurm.conf(5) man page for a full list of flags. The environment variable takes precedence over the setting in the slurm.conf. 👑
• SLURM_DELAY_BOOT — Same as --delay-boot
• SLURM_DEPENDENCY — Same as -d, --dependency=<jobid>
• SLURM_DISABLE_STATUS — Same as -X, --disable-status
• SLURM_DIST_PLANESIZE — Plane distribution size. Only used if --distribution=plane, without =<size>, is set.
• SLURM_DISTRIBUTION — Same as -m, --distribution
• SLURM_EPILOG — Same as --epilog
• SLURM_EXACT — Same as --exact
• SLURM_EXCLUSIVE — Same as --exclusive
• SLURM_EXIT_ERROR — Specifies the exit code generated when a Slurm error occurs (e.g. invalid options). This can be used by a script to distinguish application exit codes from various Slurm error conditions. Also see SLURM_EXIT_IMMEDIATE. 🎀
• SLURM_EXIT_IMMEDIATE — Specifies the exit code generated when the --immediate option is used and resources are not currently available. This can be used by a script to distinguish application exit codes from various Slurm error conditions. Also see SLURM_EXIT_ERROR.
• SLURM_EXPORT_ENV — Same as --export
• SLURM_GPU_BIND — Same as --gpu-bind
• SLURM_GPU_FREQ — Same as --gpu-freq
• SLURM_GPUS — Same as -G, --gpus
• SLURM_GPUS_PER_NODE — Same as --gpus-per-node except within an existing allocation, in which case it will be ignored if --gpus is specified.
• SLURM_GPUS_PER_TASK — Same as --gpus-per-task
• SLURM_GRES — Same as --gres. Also see SLURM_STEP_GRES
• SLURM_GRES_FLAGS — Same as --gres-flags
• SLURM_HINT — Same as --hint
• SLURM_IMMEDIATE — Same as -I, --immediate
• SLURM_JOB_ID — Same as --jobid
• SLURM_JOB_NAME — Same as -J, --job-name except within an existing allocation, in which case it is ignored to avoid using the batch job's name as the name of each job step. 🙈
• SLURM_JOB_NUM_NODES — Same as -N, --nodes. Total number of nodes in the job's resource allocation.
• SLURM_KILL_BAD_EXIT — Same as -K, --kill-on-bad-exit. Must be set to 0 or 1 to disable or enable the option.
• SLURM_LABELIO — Same as -l, --label
• SLURM_MEM_BIND — Same as --mem-bind
• SLURM_MEM_PER_CPU — Same as --mem-per-cpu
• SLURM_MEM_PER_GPU — Same as --mem-per-gpu
• SLURM_MEM_PER_NODE — Same as --mem
• SLURM_MPI_TYPE — Same as --mpi
• SLURM_NETWORK — Same as --network
• SLURM_NNODES — Same as -N, --nodes. Total number of nodes in the job's resource allocation. See SLURM_JOB_NUM_NODES. Included for backwards compatibility. 🫶
• SLURM_NO_KILL — Same as -k, --no-kill
• SLURM_NPROCS — Same as -n, --ntasks. See SLURM_NTASKS. Included for backwards compatibility.
• SLURM_NTASKS — Same as -n, --ntasks
• SLURM_NTASKS_PER_CORE — Same as --ntasks-per-core
• SLURM_NTASKS_PER_GPU — Same as --ntasks-per-gpu
• SLURM_NTASKS_PER_NODE — Same as --ntasks-per-node
• SLURM_NTASKS_PER_SOCKET — Same as --ntasks-per-socket
• SLURM_OOMKILLSTEP — Same as --oom-kill-step
• SLURM_OPEN_MODE — Same as --open-mode
• SLURM_OVERCOMMIT — Same as -O, --overcommit
• SLURM_OVERLAP — Same as --overlap
• SLURM_PARTITION — Same as -p, --partition
• SLURM_PMI_KVS_NO_DUP_KEYS — If set, then PMI key-pairs will contain no duplicate keys. MPI can use this variable to inform the PMI library that it will not use duplicate keys, so PMI can skip the check for duplicate keys. This is the case for MPICH2 and reduces overhead in testing for duplicates for improved performance. 🔥
• SLURM_POWER — Same as --power
• SLURM_PROFILE — Same as --profile
• SLURM_PROLOG — Same as --prolog
• SLURM_QOS — Same as --qos
• SLURM_REMOTE_CWD — Same as -D, --chdir=
• SLURM_REQ_SWITCH — When a tree topology is used, this defines the maximum count of switches desired for the job allocation and optionally the maximum time to wait for that number of switches. See --switches 🦋
• SLURM_RESERVATION — Same as --reservation
• SLURM_RESV_PORTS — Same as --resv-ports
• SLURM_SEND_LIBS — Same as --send-libs
• SLURM_SIGNAL — Same as --signal
• SLURM_SPREAD_JOB — Same as --spread-job
• SLURM_SRUN_REDUCE_TASK_EXIT_MSG — If set and non-zero, successive task exit messages with the same exit code will be printed only once.
• SRUN_ERROR — Same as -e, --error
• SRUN_INPUT — Same as -i, --input
• SRUN_OUTPUT — Same as -o, --output
• SLURM_STEP_GRES — Same as --gres (only applies to job steps, not to job allocations). Also see SLURM_GRES
• SLURM_STEP_KILLED_MSG_NODE_ID=ID — If set, only the specified node will log when the job or step are killed by a signal.
• SLURM_TASK_EPILOG — Same as --task-epilog
• SLURM_TASK_PROLOG — Same as --task-prolog
• SLURM_TEST_EXEC — If defined, srun will verify existence of the executable program along with user execute permission on the node where srun was called before attempting to launch it on nodes in the step. 💅
• SLURM_THREAD_SPEC — Same as --thread-spec
• SLURM_THREADS — Same as -T, --threads
• SLURM_THREADS_PER_CORE — Same as --threads-per-core
• SLURM_TIMELIMIT — Same as -t, --time ⏰
• SLURM_TRES_BIND — literally the same as --tres-bind. And heads up, bestie: if --gpu-bind is specified, it also gets set in SLURM_TRES_BIND as if you'd specified it in --tres-bind. ✨
• SLURM_TRES_PER_TASK — same as --tres-per-task. iconic behavior, no notes.
• SLURM_UMASK — if it's defined, Slurm uses that umask to set permissions when it creates the output/error files for the job. your files, your rules 💅
• SLURM_UNBUFFEREDIO — same as -u, --unbuffered.
• SLURM_USE_MIN_NODES — same as --use-min-nodes.
• SLURM_WAIT — same as -W, --wait.
• SLURM_WAIT4SWITCH — max time waiting for requested switches. peep --switches for the lore.
• SLURM_WCKEY — same as -W, --wckey.
• SLURM_WORKING_DIR — same as -D, --chdir.
• SLURMD_DEBUG — same as --slurmd-debug.
• SRUN_CONTAINER — same as --container.
• SRUN_CONTAINER_ID — same as --container-id.
• SRUN_CONTAINER_TYPE — same as --container-type.
• SRUN_EXPORT_ENV — same as --export, and it will override any setting for SLURM_EXPORT_ENV. it said what it said 🫶
• SRUN_SEGMENT_SIZE — same as --segment.

---

🌈 OUTPUT ENVIRONMENT VARIABLES 🌈

okay so this is the fun part — srun sets a whole lil ecosystem of environment variables in the environment of the executing tasks on the remote compute nodes. here's the full roster, bestie, no skips:

• SLURM_*_HET_GROUP_# — for a heterogeneous job allocation, these environment variables are set separately for each component. one per component, every time 💫
• SLURM_CLUSTER_NAME — name of the cluster on which the job is executing.
• SLURM_CPU_BIND_LIST — the --cpu-bind map or mask list, i.e. the list of Slurm CPU IDs or masks for this node, where CPU_ID = Board_ID x threads_per_board + Socket_ID x threads_per_socket + Core_ID x threads_per_core + Thread_ID. math girlie hours 🧮
• SLURM_CPU_BIND_TYPE — the --cpu-bind type (none,map_cpu:,mask_cpu:,rank_ldom,map_ldom:,mask_ldom:).
• SLURM_CPU_BIND_VERBOSE — the --cpu-bind verbosity (quiet,verbose).
• SLURM_CPU_FREQ_REQ — holds the value requested for cpu frequency on the srun command, as a numerical frequency in kilohertz, or a coded value for a request of low, medium, highm1 or high for the frequency. see the description of the --cpu-freq option or the SLURM_CPU_FREQ_REQ input environment variable. ⚡
• SLURM_CPUS_ON_NODE — number of CPUs available to the step on this node. NOTE: the select/linear plugin allocates entire nodes to jobs, so the value indicates the total count of CPUs on the node. for the cons/tres plugin, this number indicates the number of CPUs on this node allocated to the step.
• SLURM_CPUS_PER_TASK — number of cpus requested per task. only set if either the --cpus-per-task option or the --tres-per-task=cpu=# option is specified.
• SLURM_DISTRIBUTION — distribution type for the allocated jobs. set the distribution with -m, --distribution.
• SLURM_GPUS_ON_NODE — number of GPUs available to the step on this node. gpu girlies rise 🔥
• SLURM_GTIDS — global task IDs running on this node. zero origin and comma separated. it's read internally by pmi if Slurm was built with pmi support. > leaving the variable set may cause problems when using external packages from within the job (Abaqus and Ansys have been known to have problems when it is set — consult the appropriate documentation for 3rd party software). consider this your warning bestie 😭
• SLURM_HET_SIZE — set to count of components in heterogeneous job.
• SLURM_JOB_ACCOUNT — account name associated of the job allocation.
• SLURM_JOB_CPUS_PER_NODE — count of CPUs available to the job on the nodes in the allocation, using the format CPU_count[(xnumber_of_nodes)][,CPU_count[(xnumber_of_nodes)] ...]. for example, SLURM_JOB_CPUS_PER_NODE='72(x2),36' indicates that on the first and second nodes (as listed by SLURM_JOB_NODELIST) the allocation has 72 CPUs, while the third node has 36 CPUs. NOTE: the select/linear plugin allocates entire nodes to jobs, so the value indicates the total count of CPUs on allocated nodes. the select/cons_tres plugin allocates individual CPUs to jobs, so this number indicates the number of CPUs allocated to the job.
• SLURM_JOB_DEPENDENCY — set to value of the --dependency option.
• SLURM_JOB_END_TIME — the UNIX timestamp for a job's projected end time. ⏰
• SLURM_JOB_GPUS — the global GPU IDs of the GPUs allocated to this job. the GPU IDs are not relative to any device cgroup, even if devices are constrained with task/cgroup. only set in batch and interactive jobs.
• SLURM_JOB_ID — job id of the executing job. the main character 🎀
• SLURM_JOB_LICENSES — name and count of any license(s) requested.
• SLURM_JOB_NAME — set to the value of the --job-name option or the command name when srun is used to create a new job allocation. not set when srun is used only to create a job step (i.e. within an existing job allocation).
• SLURM_JOB_NODELIST — list of nodes allocated to the job.
• SLURM_JOB_NODES — total number of nodes in the job's resource allocation.
• SLURM_JOB_PARTITION — name of the partition in which the job is running.
• SLURM_JOB_QOS — Quality Of Service (QOS) of the job allocation.
• SLURM_JOB_RESERVATION — advanced reservation containing the job allocation, if any.
• SLURM_JOB_SEGMENT_SIZE — the size of the segments that was used to create the job allocation. only set if --segment is specified.
• SLURM_JOB_START_TIME — the UNIX timestamp for a job's start time.
• SLURM_JOBID — job id of the executing job. see SLURM_JOB_ID. included for backwards compatibility. the og 💜
• SLURM_LAUNCH_NODE_IPADDR — IP address of the node from which the task launch was initiated (where the srun command ran from).
• SLURM_LOCALID — node local task ID for the process within a job.
• SLURM_MEM_BIND_LIST — the --mem-bind map or mask list (<list of IDs or masks for this node>).
• SLURM_MEM_BIND_PREFER — the --mem-bind prefer (prefer).
• SLURM_MEM_BIND_TYPE — the --mem-bind type (none,rank,map_mem:,mask_mem:,local).
• SLURM_MEM_BIND_VERBOSE — the --mem-bind verbosity (quiet,verbose).
• SLURM_NETWORK — set to the value of the --network option, if specified.
• SLURM_NODEID — the relative node ID of the current node.
• SLURM_NPROCS — total number of processes in the current job or job step. see SLURM_NTASKS. included for backwards compatibility.
• SLURM_NTASKS — total number of processes in the current job or job step.
• SLURM_OVERCOMMIT — set to 1 if --overcommit was specified.
• SLURM_PRIO_PROCESS — the scheduling priority (nice value) at the time of job submission. this value is propagated to the spawned processes.
• SLURM_PROCID — the MPI rank (or relative process ID) of the current process.
• SLURM_SRUN_COMM_HOST — IP address of srun communication host.
• SLURM_SRUN_COMM_PORT — srun communication port.
• SLURM_CONTAINER — OCI Bundle for job. only set if --container is specified. 🫙
• SLURM_CONTAINER_ID — OCI id for job. only set if --container_id is specified.
• SLURM_CONTAINER_TYPE — job container type for job. only set if --container_type is specified.
• SLURM_SHARDS_ON_NODE — number of GPU Shards available to the step on this node.
• SLURM_STEP_GPUS — the global GPU IDs of the GPUs allocated to this step (excluding batch and interactive steps). the GPU IDs are not relative to any device cgroup, even if devices are constrained with task/cgroup.
• SLURM_STEP_ID — the step ID of the current job.
• SLURM_STEP_LAUNCHER_PORT — step launcher port.
• SLURM_STEP_NODELIST — list of nodes allocated to the step.
• SLURM_STEP_NUM_NODES — number of nodes allocated to the step.
• SLURM_STEP_NUM_TASKS — number of processes in the job step or whole heterogeneous job step.
• SLURM_STEP_TASKS_PER_NODE — number of processes per node within the step.
• SLURM_STEPID — the step ID of the current job. see SLURM_STEP_ID. included for backwards compatibility.
• SLURM_SUBMIT_DIR — the directory from which the allocation was invoked from.
• SLURM_SUBMIT_HOST — the hostname of the computer from which the allocation was invoked from.
• SLURM_TASK_PID — the process ID of the task being started.
• SLURM_TASKS_PER_NODE — number of tasks to be initiated on each node. values are comma separated and in the same order as SLURM_JOB_NODELIST. if two or more consecutive nodes are to have the same task count, that count is followed by "(x#)" where # is the repetition count. for example, "SLURM_TASKS_PER_NODE=2(x3),1" indicates that the first three nodes will each execute two tasks and the fourth node will execute one task. the way this format ate 💅
• SLURM_TOPOLOGY_ADDR — this is set only if the system has the topology/tree plugin configured. the value will be set to the names network switches which may be involved in the job's communications from the system's top level switch down to the leaf switch and ending with node name. a period is used to separate each hardware component name. 🦋
• SLURM_TOPOLOGY_ADDR_PATTERN — okay so this one only shows up if your system has the topology/tree plugin configured, bestie. The value is set to the component types listed in SLURM_TOPOLOGY_ADDR, and each component gets labeled either "switch" or "node". A period separates each hardware component type. 💅
• SLURM_TRES_PER_TASK — set to the value of --tres-per-task. And heads up: if --cpus-per-task or --gpus-per-task is specified, it's also set in SLURM_TRES_PER_TASK as if it were specified in --tres-per-task. Sneaky but slay.
• SLURM_UMASK — the umask that was in effect when the job was submitted. periodt.
• SLURMD_NODENAME — the name of the node running the task. And for a parallel job running across multiple compute nodes, the various tasks will have this env var set to different values on each compute node — like, each node knows its own name. iconic behavior tbh. 🌈
• SRUN_DEBUG — set to the logging level of the srun command. Default value is 3 (info level). The value gets incremented or decremented based on the --verbose and --quiet options. 🎀

---

💀 SIGNALS AND ESCAPE SEQUENCES 💀

Okay so signals sent to the srun command are automatically forwarded to the tasks it's controlling — with a few exceptions. Let's get into it, because this is the part where you really don't wanna be surprised:

• The escape sequence Ctrl+C will report the state of all tasks associated with the srun command. Cute little status check. 🫶
• If you hit Ctrl+C twice within one second, then the associated SIGINT signal will be sent to all tasks, and a termination sequence will be entered sending SIGCONT, SIGTERM, and SIGKILL to all spawned tasks. 😭 that's the "okay we're actually done here" combo.
• If a third Ctrl+C is received, the srun program will be terminated without waiting for remote tasks to exit or their I/O to complete. Yeah. It just leaves. 💀

The escape sequence Ctrl+Z is presently ignored. So don't count on it, bestie. 🙈

⚠️ Signal forwarding is not available for asynchronous steps (--async); srun exits before the step starts and holds no connection to the running tasks. Use scancel --signal to send signals to an asynchronous step. No cap, this trips people up.

---

🦋 MPI SUPPORT 🦋

MPI use depends on which MPI you're using. There are three fundamentally different modes of operation that these various MPI implementations use. Let's break it down:

1. Slurm directly launches the tasks and performs initialization of communications through the PMI2 or PMIx APIs. For example: srun -n16 a.out. Clean, simple, main character energy. ✨
2. Slurm creates a resource allocation for the job and then mpirun launches tasks using Slurm's infrastructure (OpenMPI). Teamwork, love that. 🤝
3. Slurm creates a resource allocation for the job and then mpirun launches tasks using some mechanism other than Slurm, such as SSH or RSH. These tasks are initiated outside of Slurm's monitoring or control. Slurm's epilog should be configured to purge these tasks when the job's allocation is relinquished, or the use of pam_slurm_adopt is highly recommended. 🚨 (fr fr, don't sleep on this one)

See https://slurm.schedmd.com/mpi_guide.html for more info on how to use these various MPI implementations with Slurm. Bookmark it. 📌

---

🌸 MULTIPLE PROGRAM CONFIGURATION 🌸

So, comments in the configuration file must have a "#" in column one. Column one. First character. Don't come at me about it. The configuration file contains the following fields, separated by white space:

Task rank

One or more task ranks to use this configuration. Multiple values may be comma separated. Ranges may be indicated with two numbers separated with a '-' with the smaller number first (e.g. "0-4" and not "4-0" — yes it matters, bestie, the order is the whole vibe). To indicate all tasks not otherwise specified, specify a rank of '*' as the last line of the file. And if an attempt is made to initiate a task for which no executable program is defined, you'll get this error message: "No executable program specified for this task". 😭 noted.

Executable

The name of the program to execute. May be a fully qualified pathname if desired. Go off. 💅

Arguments

Program arguments. The expression "%t" will be replaced with the task's number. The expression "%o" will be replaced with the task's offset within this range (e.g. a configured task rank value of "1-5" would have offset values of "0-4"). Single quotes may be used to avoid having the enclosed values interpreted. This field is optional. And any arguments for the program entered on the command line will be added to the arguments specified in the configuration file. 🫶

For example:

$ cat silly.conf
###################################################################
# srun multiple program configuration file
#
# srun -n8 -l --multi-prog silly.conf
###################################################################
4-6       hostname
1,7       echo  task:%t
0,2-3     echo  offset:%o

$ srun -n8 -l --multi-prog silly.conf
0: offset:0
1: task:1
2: offset:1
3: offset:2
4: linux15.llnl.gov
5: linux16.llnl.gov
6: linux17.llnl.gov
7: task:7

Look at that output. Every task doing its own little thing. Obsessed. 👑

💜 Example 1 💜

Bestie, this one's the simplest lil' flex: it runs the command hostname in eight tasks. Slurm will allocate at least eight processors to the job (same number as the task count) spread over however many nodes it takes to satisfy the request ✨. Each task's output gets prefixed with its task number, so you can tell who's who.

(The machine "dev" in the example below has a total of two CPUs per node, iykyk.)

$ srun -n8 -l hostname
0: dev0
1: dev0
2: dev1
3: dev1
4: dev2
5: dev2
6: dev3
7: dev3

---

💜 Example 2 💜

Here the srun -r option gets used inside a job script to run two job steps on disjoint nodes. The script itself is run using allocate mode instead of as a batch job in this case 🎀

$ cat test.sh
#!/bin/sh
echo $SLURM_JOB_NODELIST
srun -lN2 -r2 hostname
srun -lN2 hostname

$ salloc -N4 test.sh
dev[7-10]
0: dev9
1: dev10
0: dev7
1: dev8

---

💜 Example 3 💜

Now we're running two job steps in parallel inside an already-allocated set of nodes. Look at her go, main character energy 🔥

$ cat test.sh
#!/bin/bash
srun -lN2 -n4 -r 2 sleep 60 &
srun -lN2 -r 0 sleep 60 &
sleep 1
squeue
squeue -s
wait

$ salloc -N4 test.sh
  JOBID PARTITION     NAME     USER  ST      TIME  NODES NODELIST
  65641     batch  test.sh   grondo   R      0:01      4 dev[7-10]

STEPID     PARTITION     USER      TIME NODELIST
65641.0        batch   grondo      0:01 dev[7-8]
65641.1        batch   grondo      0:01 dev[9-10]

---

💜 Example 4 💜

This one's the classic "how do I run a simple MPI job" situation 🧚 We use srun to build a list of machines (nodes) formatted the way mpirun requires. Sample command line plus the script to execute, coming right up:

$ cat test.sh
#!/bin/sh
MACHINEFILE="nodes.$SLURM_JOB_ID"

# Generate Machinefile for mpi such that hosts are in the same
#  order as if run via srun
#
srun -l /bin/hostname | sort -n | awk '{print $2}' > $MACHINEFILE

# Run using generated Machine file:
mpirun -np $SLURM_NTASKS -machinefile $MACHINEFILE mpi-app

rm $MACHINEFILE

$ salloc -N2 -n4 test.sh

---

💜 Example 5 💜

Okay this is actually so cool — you can run different jobs on different nodes in the same srun. And you can do it for any number of nodes or any number of jobs, no cap 💅 The executables get placed on the nodes indicated by the SLURM_NODEID env var, starting at 0 and going up to the number you specified on the srun command line.

$ cat test.sh
case $SLURM_NODEID in
    0) echo "I am running on "
       hostname ;;
    1) hostname
       echo "is where I am running" ;;
esac

$ srun -N2 test.sh
dev0
is where I am running
I am running on
dev1

---

💜 Example 6 💜

Multi-core options controlling the layout of tasks 🦋 We're requesting four sockets per node and two cores per socket be dedicated to the job:

$ srun -N2 -B 4-4:2-2 a.out

---

💜 Example 7 💜

This script uses Slurm for resource management by firing off the various job steps as processors become available for their dedicated use 🍓

$ cat my.script
#!/bin/bash
srun -n4 prog1 &
srun -n3 prog2 &
srun -n1 prog3 &
srun -n1 prog4 &
wait

---

💜 Example 8 💜

Launching an application called "server" with one task, 8 CPUs and 16 GiB of memory (that's 2 GiB per CPU) — plus another application called "client" with 16 tasks, 1 CPU per task (the default) and 1 GiB of memory per task. Two apps, one srun, iconic 👑

$ srun -n1 -c8 --mem-per-cpu=2gb server : -n16 --mem-per-cpu=1gb client

---

💜 Example 9 💜

This one highlights the difference between srun's --exclusive and --overlap flags when you run from inside a job allocation 🙈 The --overlap flag lets both steps start at the same time. The --exclusive flag makes the second step wait until the first one has finished.

$ salloc  -n1
salloc: Granted job allocation 9553
salloc: Waiting for resource configuration
salloc: Nodes node01 are ready for job

$ date +%T; srun -n1 --overlap -l sleep 3 &
$ srun -n1 --overlap -l date +%T &
14:36:04
[1] 144341
[2] 144342
0: 14:36:04
[2]+  Done                    srun -n1 --overlap -l date +%T
[1]+  Done                    srun -n1 --overlap -l sleep 3

$ date +%T; srun -n1 --exclusive -l sleep 3 &
$ srun -n1 --exclusive -l date +%T &
14:36:17
[1] 144429
[2] 144430
srun: Job 9553 step creation temporarily disabled, retrying (Requested nodes are busy)
srun: Step created for job 9553
0: 14:36:20
[1]-  Done                    srun -n1 --exclusive -l sleep 3
[2]+  Done                    srun -n1 --exclusive -l date +%T

---

💜 Example 10 💜

Okay bestie, this is the "jobs that aren't evenly split across nodes" cautionary tale 💀 Here's a job that was allocated 2 CPUs on one node and 24 CPUs on the other:

$ echo $SLURM_NODELIST; echo $SLURM_JOB_CPUS_PER_NODE
node[01-02]
2,24

If a task starts that occupies the CPUs on the node with fewer CPUs, then a subsequent task that should totally be able to start on the other node will not start — because it inherits the requirement for the number of nodes from the job allocation 😭 The job step will just stay pending until the first job step completes, or until it gets cancelled.

$ srun -n4 --exact sleep 1800 &
[1] 151837

$ srun -n2 --exact hostname
^Csrun: Cancelled pending job step with signal 2
srun: error: Unable to create step for job 2677: Job/step already completing or completed

But if the job step is started while explicitly requesting a single node, then she's able to run 🫶

$ srun -n2 -N1 --exact hostname
node02
node02

You can change this behavior by adding SelectTypeParameters=CR_Pack_Nodes to your slurm.conf. That node-packing logic lets job steps start on a single node without having to explicitly request a single node.

---

💜 Example 11 💜

Last one! This demonstrates that adding the --exclusive flag to job allocation requests can give different results depending on whether you also request a certain number of tasks ⭐

Requesting exclusive access with no additional requirements lets the process access all the CPUs on the allocated node:

$ srun -l --exclusive  bash -c 'grep Cpus_allowed_list /proc/self/status'
0: Cpus_allowed_list:   0-23

Adding a request for a certain number of tasks makes each task only have access to a single CPU:

$ srun -l --exclusive -n2 bash -c 'grep Cpus_allowed_list /proc/self/status'
0: Cpus_allowed_list:   0
1: Cpus_allowed_list:   12

And you can define the number of CPUs per task if you want to give them access to more than one:

$ srun -l --exclusive -n2 --cpus-per-task=12 bash -c 'grep Cpus_allowed_list /proc/self/status'
0: Cpus_allowed_list:   0-5,12-17
1: Cpus_allowed_list:   6-11,18-23

---

🌈 COPYING 🌈

The legal bit, but make it cute — all of these facts are unchanged, promise 🎀

Copyright (C) 2006-2007 The Regents of the University of California.

Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).

Copyright (C) 2008-2010 Lawrence Livermore National Security.

Copyright (C) 2010-2022 SchedMD LLC.

This file is part of Slurm, a resource management program 💖 For details, see <https://slurm.schedmd.com/>.

Slurm is free software — you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version ✨

Slurm is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details 👀

---

🦋 SEE ALSO 🦋

Go read these next, bestie, they're all in the family:

• salloc(1)
• sattach(1)
• sbatch(1)
• sbcast(1)
• scancel(1)
• scontrol(1)
• squeue(1)
• slurm.conf(5)
• sched_setaffinity (2)
• numa (3)
• getrlimit (2)

---

Okay that's the whole page, you made it to the end 🎉💅 This was Slurm 26.05, from the Slurm Commands corner of the world, and it's all about srun(1). Go forth and launch some tasks, you absolute icon 👑🌸