TURBOSTAT(8)

💅 TURBOSTAT — the turbostat manual page (section 8, System Manager's Manual) 💖

okay bestie, sit down, hydrate, because this one is a whole vibe. this is the page for turbostat, and it's genuinely iconic, no cap.

---

✨ NAME ✨

turbostat — report processor frequency and idle statistics 🌸

---

🦋 SYNOPSIS 🦋

turbostat [Options] command
turbostat [Options] [--interval seconds]

---

👑 DESCRIPTION 👑

so like, turbostat reports processor topology, frequency, idle power-state statistics, temperature and power on X86 processors. that's the whole assignment and she ate. 🔥

there are two ways to invoke her:

1. give her a command — it gets forked, and the stats are printed one-shot when that command finishes. one and done, periodt.
2. omit the command — then turbostat just keeps displaying statistics every 5 seconds, over and over, living her main character life. you can change that 5-second interval with the --interval option. 💫

ngl, some information is just not available on older processors. not her fault, the silicon is simply not serving. 💀

🎀 Options 🎀

okay so the options situation: you can write them with a single or a double -, and you only need to type as much of the option name as it takes to disambiguate it from the others. also — and this is important — options are case-sensitive. --Dump and --dump are NOT the same girl. 🙈

• --add attributes — adds a column with a counter that has the specified attributes. the location attribute is required, everything else is optional. here's the attribute menu:
  • location: {msrDDD | msr0xXXX | /sys/path... | perf/<device>/<event>}
    • msrDDD is a decimal offset, eg. msr16
    • msr0xXXX is a hex offset, eg. msr0x10
    • /sys/path... is an absolute path to a sysfs attribute
    • <device> is a perf device from /sys/bus/event_source/devices/<device> eg. cstate_core
      • on Intel hybrid platforms, instead of one "cpu" perf device there are two: "cpu_core" and "cpu_atom" devices for P and E cores respectively.
      • turbostat, in this case, allows the user to use the "cpu" device and will automatically detect the type of a CPU and translate it to "cpu_core" and "cpu_atom" accordingly.
      • for a complete example see "ADD PERF COUNTER EXAMPLE #2 (using virtual "cpu" device)".
    • <event> is a perf event for the given device, from /sys/bus/event_source/devices/<device>/events/<event> eg. c1-residency
      • so perf/cstate_core/c1-residency would then use /sys/bus/event_source/devices/cstate_core/events/c1-residency ✨
  • scope: {cpu | core | package}
    • sample and print the counter for every cpu, core, or package.
    • default: cpu
  • size: {u32 | u64}
    • MSRs are read as 64-bits, u32 truncates the displayed value to 32-bits.
    • default: u64
  • format: {raw | delta | percent | average}
    • 'raw' shows the MSR contents in hex.
    • 'delta' shows the difference in values during the measurement interval.
    • 'percent' shows the delta as a percentage of the cycles elapsed.
    • 'average' similar to raw, but also averaged for node/package summaries (or when using -S).
    • default: delta
  • name: "name_string"
    • any string that does not match a key-word above is used as the column header.
• --add pmt,[attr_name=attr_value, ...] — adds a column with a PMT (Intel Platform Monitoring Technology) counter in a similar way to the --add option above, but it requires PMT metadata to be supplied to correctly read and display the counter. the metadata can be found in the Intel PMT XML files, hosted at https://github.com/intel/Intel-PMT. for a complete example see "ADD PMT COUNTER EXAMPLE". 🌈
  • name="name_string"
    • for column header.
  • type={raw}
    • 'raw' shows the counter contents in hex.
    • default: raw
  • format={raw | delta}
    • 'raw' shows the counter contents in hex.
    • 'delta' shows the difference in values during the measurement interval.
    • default: raw
  • domain={cpu%u | core%u | package%u}
    • 'cpu' per cpu/thread counter.
    • 'core' per core counter.
    • 'package' per package counter.
    • '%u' denotes id of the domain that the counter is associated with. for example core4 would mean that the counter is associated with core number 4.
  • offset=%u
    • '%u' offset within the PMT MMIO region.
  • lsb=%u
    • '%u' least significant bit within the 64 bit value read from 'offset'. together with 'msb', used to form a read mask.
  • msb=%u
    • '%u' most significant bit within the 64 bit value read from 'offset'. together with 'lsb', used to form a read mask.
  • guid=%x
    • '%x' hex identifier of the PMT MMIO region.
• --cpu cpu-set — limits output to the system summary plus the specified cpu-set. if cpu-set is the string "core", then the system summary plus the first CPU in each core are printed — eg. subsequent HT siblings are not printed. or if cpu-set is the string "package", then the system summary plus the first CPU in each package is printed. otherwise, the system summary plus the specified set of CPUs are printed. the cpu-set is ordered from low to high, comma delimited with ".." and "-" permitted to denote a range. eg. 1,2,8,14..17,21-44 🎯
• --hide column — do not show the specified built-in columns. may be invoked multiple times, or with a comma-separated list of column names.
• --enable column — show the specified built-in columns, which are otherwise disabled by default. currently the only built-in counters disabled by default are "usec", "Time_Of_Day_Seconds", "APIC" and "X2APIC". the column name "all" can be used to enable all disabled-by-default built-in counters. 💅
• --show column — show only the specified built-in columns. may be invoked multiple times, or with a comma-separated list of column names.
• --show CATEGORY --hide CATEGORY — show and hide also accept a comma-separated-list of CATEGORIES of columns: "all", "topology", "idle", "frequency", "power", "cpuidle", "hwidle", "swidle", "cache", "llc", "other".
  • "idle" (enabled by default), includes "hwidle" and "pct_idle".
  • "cpuidle" (default disabled) includes cpuidle software invocation counters.
  • "swidle" includes "cpuidle" plus "pct_idle".
  • "hwidle" includes only hardware based idle residency counters.
  • older versions of turbostat used the term "sysfs" for what is now "swidle". history, bestie. 📜
• --Dump — displays the raw counter values. 👀
• --quiet — do not decode and print the system configuration header information. she's being mysterious.
• --no-msr — disable all the uses of the MSR driver.
• --no-perf — disable all the uses of the perf API.
• --force — force turbostat to run on an unsupported platform (minimal defaults). unsupported platform energy, she's being brave. 😭
• --interval seconds — overrides the default 5.0 second measurement interval. ⏰
• --num_iterations num — number of the measurement iterations.
• --header_iterations num — print header every num iterations.
• --out output_file — turbostat output is written to the specified output_file. the file is truncated if it already exists, and it is created if it does not exist. heads up bestie, she will NOT hesitate to overwrite. 💀
• --help — displays usage for the most common parameters.
• --Joules — displays energy in Joules, rather than dividing Joules by time to print power in Watts. ⚡
• --list — display column header names available for use by --show and --hide, then exit.
• --Summary — limits output to a 1-line System Summary for each interval.
• --TCC temperature — sets the Thermal Control Circuit temperature for systems which do not export that value. this is used for making sense of the Digital Thermal Sensor outputs, as they return degrees Celsius below the TCC activation temperature. 🌡️
• --version — displays the version.
• the command parameter — forks command, and upon its exit, displays the statistics gathered since it was forked. 🫶

---

🌈 ROW DESCRIPTIONS 🌈

the system configuration dump (if --quiet is not used) is followed by statistics. here's how to read the rows, bestie:

• the first row of the statistics labels the content of each column (below).
• the second row of statistics is the system summary line. 👑 the system summary line has a '-' in the columns for the Package, Core, and CPU.
• the contents of the system summary line depends on the type of column:
  • columns that count items (eg. IRQ) show the sum across all CPUs in the system.
  • columns that show a percentage show the average across all CPUs in the system.
  • columns that dump raw MSR values simply show 0 in the summary. 💀
• after the system summary row, each row describes a specific Package/Core/CPU.

⚠️ note that if the --cpu parameter is used to limit which specific CPUs are displayed, turbostat will still collect statistics for all CPUs in the system and will still show the system summary for all CPUs in the system. she sees everything. 👁️✨

🌈 COLUMN DESCRIPTIONS 💫

okay bestie, this is the part where we go column by column, and there are so many of them, but every single one is here, no skips, no cap. grab a snack. 🍓

---

• usec — for each CPU, the number of microseconds that elapsed while counters were being collected, thread migration included (if any happened). This one is disabled by default, and you turn it on with "--enable usec", or with --debug. On the summary row, usec means the total elapsed time it took to snapshot the procfs/sysfs and collect the counters across all cpus.
• Time_Of_Day_Seconds — for each CPU, the gettimeofday(2) value (seconds.subsec since Epoch) for when the counters that closed out the measurement interval got collected. Disabled by default, enable it with "--enable Time_Of_Day_Seconds" or "--debug". On the summary row, Time_Of_Day_Seconds is the timestamp right after counters were collected on the last CPU.
• Core — the processor core number. bestie, if you see multiple CPUs hanging out under one core, that's support for Intel(R) Hyper-Threading Technology (HT). iconic behavior tbh.
• CPU — the Linux CPU (logical processor) number. And yes, it is completely fine that on lots of systems the CPUs are not listed in numerical order — for efficiency reasons, turbostat runs in topology order, so HT siblings show up together. it's giving organized. 💅
• Package — the processor package number — not present on systems that only have a single processor package.
• Avg_MHz — number of cycles executed divided by time elapsed. heads up: this includes idle-time, when 0 instructions are executed.
• Busy% — the percent of the measurement interval that the CPU is executing instructions, aka. % of time spent in "C0" state.
• Bzy_MHz — the average clock rate while the CPU was not idle (ie. while it was in "c0" state).
• TSC_MHz — the average MHz that the TSC ran during the entire interval.
• IRQ — the number of interrupts serviced by that CPU during the measurement interval. The system total line is the sum of interrupts serviced across all CPUs. turbostat parses /proc/interrupts to generate this summary.
• SMI — the number of System Management Interrupts serviced CPU during the measurement interval. Even though this counter is technically per-CPU, SMI are triggered on all processors, so the number should be the same for all CPUs.
• LLCMRPS — Last Level Cache Millions of References Per Second. For CPUs with an L3 LLC, this is the number of references that CPU made to the L3 (plus the number of misses that CPU made to its L2). For CPUs with an L2 LLC, this is the number of references to the L2 (plus the number of misses to the CPU's L1). The system summary row shows the sum for all CPUs. Either way, the value displayed is the actual value divided by 1,000,000. If this value is large, then the LLC%hit column is significant. If it's small, then the LLC%hit column is not significant.
• LLC%hit — Last Level Cache Hit Rate %. The formula is Hit Rate Percent = 100.0 * Hits/References. The system summary row shows the weighted average for all CPUs (100.0 * Sum_Hits/Sum_References).
• L2MRPS — Level-2 Cache Millions of References Per Second. For CPUs with an L2 LLC, this is the same as LLC references. The system summary row shows the sum for all CPUs. Either way, the value displayed is the actual value divided by 1,000,000. If this value is large, then the L2%hit column is significant. If it's small, then the L2%hit column is not significant.
• L2%hit — Level-2 Cache Hit Rate %. Hit Rate Percent = 100.0 * Hits/References. The system summary row shows the weighted average for all CPUs (100.0 * (Sum_Hits)/Sum_References).
• C1, C2, C3... — the number of times Linux requested the C1, C2, C3 idle state during the measurement interval. The system summary line shows the sum for all CPUs. These are C-state names as exported in /sys/devices/system/cpu/cpu*/cpuidle/state*/name. Their names are generic, but their attributes are processor specific. The system description section of output shows what MWAIT sub-states they map to on each system. These counters live in the "cpuidle" group, which is disabled by default.
• C1+, C2+, C3+... — the idle governor idle state misprediction statistics. This shows the number of times Linux requested the C1, C2, C3 idle state during the measurement interval, but should have requested a deeper idle state (if it exists and is enabled). These stats come from the /sys/devices/system/cpu/cpu*/cpuidle/state*/below file. These counters are in the "cpuidle" group, which is disabled by default.
• C1-, C2-, C3-... — the idle governor idle state misprediction statistics, but the other direction. This shows the number of times Linux requested the C1, C2, C3 idle state during the measurement interval, but should have requested a shallower idle state (if it exists and is enabled). These stats come from the /sys/devices/system/cpu/cpu*/cpuidle/state*/above file. These counters are in the "cpuidle" group, which is disabled by default.
• C1%, C2%, C3% — the residency percentage that Linux requested C1, C2, C3.... The system summary is the average of all CPUs in the system. note bestie: these are software, reflecting what was requested. The hardware counters reflect what was actually achieved. These counters are in the "pct_idle" group, which is enabled by default.
• CPU%c1, CPU%c3, CPU%c6, CPU%c7 — show the percentage residency in hardware core idle states. These numbers come from hardware residency counters and live in the "hwidle" group, which is enabled by default.
• CoreTmp — degrees Celsius reported by the per-core Digital Thermal Sensor.
• PkgTmp — degrees Celsius reported by the per-package Package Thermal Monitor.
• CoreThr — Core Thermal Throttling events during the measurement interval. btw, events since boot can be found in /sys/devices/system/cpu/cpu*/thermal_throttle/*.
• GFX%rc6 — the percentage of time the GPU is in the "render C6" state, rc6, during the measurement interval. It comes from /sys/class/drm/card0/power/rc6_residency_ms or /sys/class/drm/card0/gt/gt0/rc6_residency_ms or /sys/class/drm/card0/device/tile0/gtN/gtidle/idle_residency_ms, depending on which graphics driver is being used.
• GFXMHz — an instantaneous snapshot of what sysfs presents at the end of the measurement interval. From /sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz or /sys/class/drm/card0/gt_cur_freq_mhz or /sys/class/drm/card0/gt/gt0/rps_cur_freq_mhz or /sys/class/drm/card0/device/tile0/gtN/freq0/cur_freq, depending on which graphics driver is being used.
• GFXAMHz — another instantaneous snapshot of what sysfs presents at the end of the measurement interval. From /sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz or /sys/class/drm/card0/gt_act_freq_mhz or /sys/class/drm/card0/gt/gt0/rps_act_freq_mhz or /sys/class/drm/card0/device/tile0/gtN/freq0/act_freq, depending on which graphics driver is being used.
• SAM%mc6 — the percentage of time the SA Media is in the "module C6" state, mc6, during the measurement interval. From /sys/class/drm/card0/gt/gt1/rc6_residency_ms or /sys/class/drm/card0/device/tile0/gtN/gtidle/idle_residency_ms, depending on which graphics driver is being used.
• SAMMHz — an instantaneous snapshot of what sysfs presents at the end of the measurement interval. From /sys/class/drm/card0/gt/gt1/rps_cur_freq_mhz or /sys/class/drm/card0/device/tile0/gtN/freq0/cur_freq, depending on which graphics driver is being used.
• SAMAMHz — an instantaneous snapshot of what sysfs presents at the end of the measurement interval. From /sys/class/drm/card0/gt/gt1/rps_act_freq_mhz or /sys/class/drm/card0/device/tile0/gtN/freq0/act_freq, depending on which graphics driver is being used.
• Totl%C0 — weighted percentage of time that CPUs are busy. If N CPUs are busy during an interval, the percentage is N * 100%.
• Any%C0 — percentage of time that at least one CPU is busy.
• GFX%C0 — percentage of time that at least one GFX compute engine is busy.
• CPUGFX% — percentage of time that at least one CPU is busy at the same time as at least one Graphics compute engine is busy. multitasking queens. 👑
• Pkg%pc2, Pkg%pc3, Pkg%pc6, Pkg%pc7 — percentage residency in hardware package idle states. These numbers are from hardware residency counters.
• PkgWatt — Watts consumed by the whole package.
• CorWatt — Watts consumed by the core part of the package.
• GFXWatt — Watts consumed by the Graphics part of the package — available only on client processors.
• RAMWatt — Watts consumed by the DRAM DIMMS — available only on server processors.
• SysWatt — Watts consumed by the whole platform (RAPL PSYS).
• PKG_% — percent of the interval that RAPL throttling was active on the Package. Heads up: the system summary is the sum of the package throttling time, so on a multi-package system it may be higher than 100%. Also note that the meaning of this field is model specific. For example, some hardware increments this counter when RAPL responds to thermal limits, but does not increment this counter when RAPL responds to power limits. Comparing PkgWatt and PkgTmp to system limits is necessary.
• RAM_% — percent of the interval that RAPL throttling was active on DRAM.
• UncMHz — per-package uncore MHz, instantaneous sample.
• UMHz1.0 — per-package uncore MHz for pm_domain=1 and fabric_cluster=0, instantaneous sample. The system summary is the average of all packages. Intel Granite Rapids systems use pm_domains 0-2 for CPUs, and 3-4 for IO, with cluster always 0. For the "--show" and "--hide" options, use "UncMHz" to operate on all UMHz*.* as a group.

😵‍💫 TOO MUCH INFORMATION EXAMPLE 💀

okay so by default, turbostat is that one friend who tells you literally everything. It dumps all possible information — a system configuration header, and then columns for every single counter. And honestly? That's iconic for remote debugging. Just use the --out option to save the whole thing to a text file, and then get that file to the expert who's helping you debug. They'll be obsessed with you. 🫶

But when you are not interested in all that information, bestie, there are several ways to see only what you want:

• First, the --quiet option will skip the configuration information, and turbostat will show only the counter columns. Minimal. Clean. Slay. ✨
• Second, you can reduce the columns with the --hide and --show options. Use --show and turbostat will show only the columns you list. Use --hide and turbostat will show all the columns, except the ones you list. Pick your fighter. 💅

To find out what columns are available for --show and --hide, the --list option is right there for you. Usually, the CATEGORY names above are what you use to refer to groups of counters. Also, for convenience, the special string "sysfs" can be used to refer to all of the sysfs C-state counters at once:

sudo ./turbostat --show sysfs --quiet sleep 10
10.003837 sec
    C1   C1E  C3   C6   C7s  C1%  C1E% C3%  C6%  C7s%
    4    21   2    2    459  0.14 0.82 0.00 0.00 98.93
    1    17   2    2    130  0.00 0.02 0.00 0.00 99.80
    0    0    0    0    31   0.00 0.00 0.00 0.00 99.95
    2    1    0    0    52   1.14 6.49 0.00 0.00 92.21
    1    2    0    0    52   0.00 0.08 0.00 0.00 99.86
    0    0    0    0    71   0.00 0.00 0.00 0.00 99.89
    0    0    0    0    25   0.00 0.00 0.00 0.00 99.96
    0    0    0    0    74   0.00 0.00 0.00 0.00 99.94
    0    1    0    0    24   0.00 0.00 0.00 0.00 99.84

---

🎯 ONE SHOT COMMAND EXAMPLE

So here's the tea: if turbostat is invoked with a command, it will fork that command and output the statistics it gathered after the command exits. In this case, turbostat output goes to stderr, by default. (Not stdout! stderr! don't @ me.) But you can instead save the output to a file using the --out option. 💾

In this example, the "sleep 10" command is forked, and turbostat waits for it to complete before saving all the statistics into "ts.out". Note that "sleep 10" is not part of turbostat — it's just a little example of a command that turbostat can fork. And that "ts.out" file? That's exactly what you want to open in a very wide window, paste into a spreadsheet, or attach to a bugzilla entry. main character file behaviour, fr fr.

[root@hsw]# ./turbostat -o ts.out sleep 10
[root@hsw]#

---

⏰ PERIODIC INTERVAL EXAMPLE

Without a command to fork, turbostat just vibes and displays statistics every 5 seconds. Periodic output goes to stdout, by default, unless --out is used to specify an output file. And that 5-second interval? You can change it with the -i sec option. 🦋

sudo turbostat --quiet --show CPU,frequency
     Core CPU  Avg_MHz   Busy%     Bzy_MHz   TSC_MHz   CPU%c7    UncMhz
     -    -    524  12.48     4198 3096 74.53     3800
     0    0    4    0.09 4081 3096 98.88     3800
     0    4    1    0.02 4063 3096
     1    1    2    0.06 4063 3096 99.60
     1    5    2    0.05 4070 3096
     2    2    4178 99.52     4199 3096 0.00
     2    6    3    0.08 4159 3096
     3    3    1    0.04 4046 3096 99.66
     3    7    0    0.01 3989 3096
     Core CPU  Avg_MHz   Busy%     Bzy_MHz   TSC_MHz   CPU%c7    UncMhz
     -    -    525  12.52     4198 3096 74.54     3800
     0    0    4    0.10 4051 3096 99.49     3800
     0    4    2    0.04 3993 3096
     1    1    3    0.07 4054 3096 99.56
     1    5    4    0.10 4018 3096
     2    2    4178 99.51     4199 3096 0.00
     2    6    4    0.09 4143 3096
     3    3    2    0.06 4026 3096 99.10
     3    7    7    0.17 4074 3096

This example also shows the use of the --show option to show only the desired columns. Curated. ✨

---

🖥️ SYSTEM CONFIGURATION INFORMATION EXAMPLE

By default, turbostat always dumps system configuration information before it starts taking measurements. In the example above, "--quiet" is used to suppress that output. But here's what that configuration information actually looks like, in all its glory:

turbostat version 2022.04.16 - Len Brown <lenb@kernel.org>
Kernel command line: BOOT_IMAGE=/boot/vmlinuz-5.18.0-rc6-00001-ge6891250e3b5 ...
CPUID(0): GenuineIntel 0x16 CPUID levels
CPUID(1): family:model:stepping 0x6:9e:9 (6:158:9) microcode 0xea
CPUID(0x80000000): max_extended_levels: 0x80000008
CPUID(1): SSE3 MONITOR - EIST TM2 TSC MSR ACPI-TM HT TM
CPUID(6): APERF, TURBO, DTS, PTM, HWP, HWPnotify, HWPwindow, HWPepp, No-HWPpkg, EPB
cpu7: MSR_IA32_MISC_ENABLE: 0x00850089 (TCC EIST MWAIT PREFETCH TURBO)
CPUID(7): SGX
cpu7: MSR_IA32_FEATURE_CONTROL: 0x00000005 (Locked )
CPUID(0x15): eax_crystal: 2 ebx_tsc: 258 ecx_crystal_hz: 0
TSC: 3096 MHz (24000000 Hz * 258 / 2 / 1000000)
CPUID(0x16): base_mhz: 3100 max_mhz: 4200 bus_mhz: 100
cpu7: MSR_MISC_PWR_MGMT: 0x00401cc0 (ENable-EIST_Coordination DISable-EPB DISable-OOB)
RAPL: 5825 sec. Joule Counter Range, at 45 Watts
cpu7: MSR_PLATFORM_INFO: 0x80839f1011f00
8 * 100.0 = 800.0 MHz max efficiency frequency
31 * 100.0 = 3100.0 MHz base frequency
cpu7: MSR_IA32_POWER_CTL: 0x002c005d (C1E auto-promotion: DISabled)
cpu7: MSR_TURBO_RATIO_LIMIT: 0x2728292a
39 * 100.0 = 3900.0 MHz max turbo 4 active cores
40 * 100.0 = 4000.0 MHz max turbo 3 active cores
41 * 100.0 = 4100.0 MHz max turbo 2 active cores
42 * 100.0 = 4200.0 MHz max turbo 1 active cores
cpu7: MSR_CONFIG_TDP_NOMINAL: 0x0000001f (base_ratio=31)
cpu7: MSR_CONFIG_TDP_LEVEL_1: 0x00000000 ()
cpu7: MSR_CONFIG_TDP_LEVEL_2: 0x00000000 ()
cpu7: MSR_CONFIG_TDP_CONTROL: 0x80000000 ( lock=1)
cpu7: MSR_TURBO_ACTIVATION_RATIO: 0x00000000 (MAX_NON_TURBO_RATIO=0 lock=0)
cpu7: MSR_PKG_CST_CONFIG_CONTROL: 0x1e008008 (UNdemote-C3, UNdemote-C1, demote-C3, demote-C1, locked, pkg-cstate-limit=8 (unlimited))
Uncore Frequency pkg0 die0: 800 - 3900 MHz (800 - 3900 MHz)
/dev/cpu_dma_latency: 2000000000 usec (default)
current_driver: intel_idle
current_governor: menu
current_governor_ro: menu
cpu7: POLL: CPUIDLE CORE POLL IDLE
cpu7: C1: MWAIT 0x00
cpu7: C1E: MWAIT 0x01
cpu7: C3: MWAIT 0x10
cpu7: C6: MWAIT 0x20
cpu7: C7s: MWAIT 0x33
cpu7: C8: MWAIT 0x40
cpu7: C9: MWAIT 0x50
cpu7: C10: MWAIT 0x60
cpu7: cpufreq driver: intel_pstate
cpu7: cpufreq governor: performance
cpufreq intel_pstate no_turbo: 0
cpu7: MSR_MISC_FEATURE_CONTROL: 0x00000000 (L2-Prefetch L2-Prefetch-pair L1-Prefetch L1-IP-Prefetch)
cpu0: MSR_PM_ENABLE: 0x00000001 (HWP)
cpu0: MSR_HWP_CAPABILITIES: 0x01101f53 (high 83 guar 31 eff 16 low 1)
cpu0: MSR_HWP_REQUEST: 0x00005353 (min 83 max 83 des 0 epp 0x0 window 0x0 pkg 0x0)
cpu0: MSR_HWP_INTERRUPT: 0x00000001 (EN_Guaranteed_Perf_Change, Dis_Excursion_Min)
cpu0: MSR_HWP_STATUS: 0x00000004 (No-Guaranteed_Perf_Change, No-Excursion_Min)
cpu0: EPB: 6 (balanced)
cpu0: MSR_RAPL_POWER_UNIT: 0x000a0e03 (0.125000 Watts, 0.000061 Joules, 0.000977 sec.)
cpu0: MSR_PKG_POWER_INFO: 0x00000168 (45 W TDP, RAPL 0 - 0 W, 0.000000 sec.)
cpu0: MSR_PKG_POWER_LIMIT: 0x42820800218208 (UNlocked)
cpu0: PKG Limit #1: ENabled (65.000 Watts, 64.000000 sec, clamp ENabled)
cpu0: PKG Limit #2: ENabled (65.000 Watts, 0.002441* sec, clamp DISabled)
cpu0: MSR_VR_CURRENT_CONFIG: 0x00000000
cpu0: PKG Limit #4: 0.000000 Watts (UNlocked)
cpu0: MSR_DRAM_POWER_LIMIT: 0x5400de00000000 (UNlocked)
cpu0: DRAM Limit: DISabled (0.000 Watts, 0.000977 sec, clamp DISabled)
cpu0: MSR_PP0_POLICY: 0
cpu0: MSR_PP0_POWER_LIMIT: 0x00000000 (UNlocked)
cpu0: Cores Limit: DISabled (0.000 Watts, 0.000977 sec, clamp DISabled)
cpu0: MSR_PP1_POLICY: 0
cpu0: MSR_PP1_POWER_LIMIT: 0x00000000 (UNlocked)
cpu0: GFX Limit: DISabled (0.000 Watts, 0.000977 sec, clamp DISabled)
cpu0: MSR_IA32_TEMPERATURE_TARGET: 0x00640000 (100 C) (100 default - 0 offset)
cpu0: MSR_IA32_PACKAGE_THERM_STATUS: 0x88200800 (68 C)
cpu0: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x00000003 (100 C, 100 C)
cpu7: MSR_PKGC3_IRTL: 0x0000884e (valid, 79872 ns)
cpu7: MSR_PKGC6_IRTL: 0x00008876 (valid, 120832 ns)
cpu7: MSR_PKGC7_IRTL: 0x00008894 (valid, 151552 ns)
cpu7: MSR_PKGC8_IRTL: 0x000088fa (valid, 256000 ns)
cpu7: MSR_PKGC9_IRTL: 0x0000894c (valid, 339968 ns)
cpu7: MSR_PKGC10_IRTL: 0x00008bf2 (valid, 1034240 ns)

okay let's decode this a lil bit, because it's a lot. 🫠

The max efficiency frequency, a.k.a. Low Frequency Mode, is the frequency available at the minimum package voltage. The TSC frequency is the base frequency of the processor — this should match the brand string in /proc/cpuinfo. And this base frequency should be sustainable on all CPUs indefinitely, given nominal power and cooling. (Sustainable! Indefinitely! she's committed. 💅) The remaining rows show what maximum turbo frequency is possible depending on the number of idle cores.

⚠️ Note that not all information is available on all processors. So if your output looks a little different, that's not you doing it wrong, bestie.

---

➕ ADD COUNTER EXAMPLE

Here we limit turbostat to showing just the CPU number for cpu0 - cpu3. Then we add a counter showing the 32-bit raw value of MSR 0x199 (MSR_IA32_PERF_CTL), labeling it with the column header, "PRF_CTRL", and display it only once, after the conclusion of a 0.1 second sleep. So extra, so cute. 🎀

sudo ./turbostat --quiet --cpu 0-3 --show CPU --add msr0x199,u32,raw,PRF_CTRL sleep .1
0.101604 sec
CPU    PRF_CTRL
-    0x00000000
0    0x00000c00
1    0x00000800
2    0x00000a00
3    0x00000800

---

➕✨ ADD PERF COUNTER EXAMPLE

Same energy as before: we limit turbostat to showing just the CPU number for cpu0 - cpu3. This time we add a counter showing time spent in C1 core cstate, labeling it with the column header, "pCPU%c1", and display it only once, after the conclusion of a 0.1 second sleep. We also show the CPU%c1 built-in counter, which should show similar values. (A little compare-and-contrast moment, we love to see it. 👑)

sudo ./turbostat --quiet --cpu 0-3 --show CPU,CPU%c1 --add perf/cstate_core/c1-residency,cpu,delta,percent,pCPU%c1 sleep .1
0.102448 sec
CPU     pCPU%c1 CPU%c1
-       34.89   34.89
0       45.99   45.99
1       45.94   45.94
2       23.83   23.83
3       23.84   23.84

ADD PERF COUNTER EXAMPLE #2 (using virtual cpu device) ✨🦋

Okay so picture this: you're on a hybrid, Meteor Lake platform. We're limiting turbostat to only show output for cpu0 (pcore) and cpu4 (ecore). We add a counter for number of L3 cache misses using virtual "cpu" device, and label it with the column header "VCMISS". Then another L3 cache misses counter using virtual "cpu_core" device, labeled "PCMISS". Heads up: this will fail on ecore cpu4. Then another L3 cache misses counter using virtual "cpu_atom" device, labeled "ECMISS". This will fail on pcore cpu0. We display it only once, after the conclusion of 0.1 second sleep.

sudo ./turbostat --quiet --cpu 0,4 --show CPU --add perf/cpu/cache-misses,cpu,delta,VCMISS --add perf/cpu_core/cache-misses,cpu,delta,PCMISS --add perf/cpu_atom/cache-misses,cpu,delta,ECMISS sleep 5
turbostat: added_perf_counters_init_: perf/cpu_atom/cache-misses: failed to open counter on cpu0
turbostat: added_perf_counters_init_: perf/cpu_core/cache-misses: failed to open counter on cpu4
5.001207 sec
CPU  ECMISS         PCMISS         VCMISS
-    41586506  46291219  87877749
4    83173012  0         83173040
0    0         92582439  92582458

ADD PMT COUNTER EXAMPLE 💅

Here we limit turbostat to showing just the CPU number 0. We add two counters, showing crystal clock count and the DC6 residency. All the parameters passed are based on the metadata found in the PMT XML files.

For the crystal clock count:

• We label it with the column header "XTAL".
• We set the type to 'raw', to read the number of clock ticks in hex.
• We set the format to 'delta', to display the difference in ticks during the measurement interval.
• We set the domain to 'package0', to collect it and associate it with the whole package number 0.
• We set the offset to '0', which is a offset of the counter within the PMT MMIO region.
• We set the lsb and msb to cover all 64 bits of the read 64 bit value.
• And finally we set the guid to '0x1a067102', that identifies the PMT MMIO region to which the 'offset' is applied to read the counter value.

For the DC6 residency counter:

• We label it with the column header "Die%c6".
• We set the type to 'txtal_time'.
• To obtain the percent residency value we set the format to 'delta', to display the difference in ticks during the measurement interval.
• We set the domain to 'package0', to collect it and associate it with the whole package number 0.
• We set the offset to '0', which is a offset of the counter within the PMT MMIO region.
• We set the lsb and msb to cover all 64 bits of the read 64 bit value.
• And finally we set the guid to '0x1a067102', that identifies the PMT MMIO region to which the 'offset' is applied to read the counter value.

sudo ./turbostat --quiet --cpu 0 --show CPU --add pmt,name=XTAL,type=raw,format=delta,domain=package0,offset=0,lsb=0,msb=63,guid=0x1a067102 --add pmt,name=Die%c6,type=txtal_time,format=delta,domain=package0,offset=120,lsb=0,msb=63,guid=0x1a067102
0.104352 sec
CPU                   XTAL      Die%c6
-       0x0000006d4d957ca7      0.00
0       0x0000006d4d957ca7      0.00
0.102448 sec

INPUT 📥

For interval-mode, turbostat will immediately end the current interval when it sees a newline on standard input. turbostat will then start the next interval. Control-C will send a SIGINT to turbostat, which will immediately abort the program with no further processing.

SIGNALS 🚨

SIGINT will interrupt interval-mode. The end-of-interval data will be collected and displayed before turbostat exits.

SIGUSR1 will end current interval, end-of-interval data will be collected and displayed before turbostat starts a new interval.

NOTES 📝

turbostat must be run as root. Alternatively, non-root users can be enabled to run turbostat this way:

# setcap cap_sys_admin,cap_sys_rawio,cap_sys_nice=+ep path/to/turbostat

# chmod +r /dev/cpu/*/msr

# chmod +r /dev/cpu_dma_latency

turbostat reads hardware counters, but doesn't write them. So it will not interfere with the OS or other programs, including multiple invocations of itself.

turbostat may work poorly on Linux-2.6.20 through 2.6.29, as acpi-cpufreq periodically cleared the APERF and MPERF MSRs in those kernels.

AVG_MHz = APERF_delta/measurement_interval. This is the actual number of elapsed cycles divided by the entire sample interval -- including idle time. Note that this calculation is resilient to systems lacking a non-stop TSC.

TSC_MHz = TSC_delta/measurement_interval. On a system with an invariant TSC, this value will be constant and will closely match the base frequency value shown in the brand string in /proc/cpuinfo. On a system where the TSC stops in idle, TSC_MHz will drop below the processor's base frequency.

Busy% = MPERF_delta/TSC_delta

Bzy_MHz = TSC_delta*APERF_delta/MPERF_delta/measurement_interval

Note that these calculations depend on TSC_delta, so they are not reliable during intervals when TSC_MHz is not running at the base frequency.

Turbostat data collection is not atomic. Extremely short measurement intervals (much less than 1 second), or system activity that prevents turbostat from being able to run on all CPUS to quickly collect data, will result in inconsistent results.

The APERF, MPERF MSRs are defined to count non-halted cycles. Although it is not guaranteed by the architecture, turbostat assumes that they count at TSC rate, which is true on all processors tested to date.

REFERENCES 📚

Volume 3B: System Programming Guide"

https://www.intel.com/products/processor/manuals/

RUN THE LATEST VERSION 🚀

If turbostat complains that it doesn't recognize your processor, please try the latest version.

The latest version of turbostat does not require the latest version of the Linux kernel. However, some features, such as perf(1) counters, do require kernel support.

The latest turbostat release is available in the upstream Linux Kernel source tree. eg. "git pull https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git" and run make in tools/power/x86/turbostat/.

n.b. "make install" will update your system manually, but a distro update may subsequently downgrade your turbostat to an older version. For this reason, manually installing to /usr/local/bin may be what you want.

Note that turbostat/Makefile has a "make snapshot" target, which will create a tar file that can build without a local kernel source tree.

If the upstream version isn't new enough, the development tree can be found here: "git pull https://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux.git turbostat"

If the development tree doesn't work, please contact the author via chat, or via email with the word "turbostat" on the Subject line.

An old turbostat binary may run on unknown hardware by using "--force", but results are unsupported.

FILES 🗂️

• /sys/bus/event_source/devices/
• /dev/cpu/*/msr
• /sys/class/intel_pmt/
• /sys/devices/system/cpu/

SEE ALSO 👀

perf(1), msr(4), vmstat(8)

AUTHOR 💖

Written by Len Brown <len.brown@intel.com>

---

That's the whole TURBOSTAT(8) manual page, bestie — source/version and date live in your local footer, but this vibe is forever. Now go be the main character of your own perf counter 💅✨💖