xref: /linux/tools/perf/Documentation/perf-stat.txt (revision 3ea5eb68b9d624935108b5e696859304edfac202)
1perf-stat(1)
2============
3
4NAME
5----
6perf-stat - Run a command and gather performance counter statistics
7
8SYNOPSIS
9--------
10[verse]
11'perf stat' [-e <EVENT> | --event=EVENT] [-a] <command>
12'perf stat' [-e <EVENT> | --event=EVENT] [-a] \-- <command> [<options>]
13'perf stat' [-e <EVENT> | --event=EVENT] [-a] record [-o file] \-- <command> [<options>]
14'perf stat' report [-i file]
15
16DESCRIPTION
17-----------
18This command runs a command and gathers performance counter statistics
19from it.
20
21
22OPTIONS
23-------
24<command>...::
25	Any command you can specify in a shell.
26
27record::
28	See STAT RECORD.
29
30report::
31	See STAT REPORT.
32
33-e::
34--event=::
35	Select the PMU event. Selection can be:
36
37	- a symbolic event name (use 'perf list' to list all events)
38
39	- a raw PMU event in the form of rN where N is a hexadecimal value
40	  that represents the raw register encoding with the layout of the
41	  event control registers as described by entries in
42	  /sys/bus/event_source/devices/cpu/format/*.
43
44        - a symbolic or raw PMU event followed by an optional colon
45	  and a list of event modifiers, e.g., cpu-cycles:p.  See the
46	  linkperf:perf-list[1] man page for details on event modifiers.
47
48	- a symbolically formed event like 'pmu/param1=0x3,param2/' where
49	  param1 and param2 are defined as formats for the PMU in
50	  /sys/bus/event_source/devices/<pmu>/format/*
51
52	  'percore' is a event qualifier that sums up the event counts for both
53	  hardware threads in a core. For example:
54	  perf stat -A -a -e cpu/event,percore=1/,otherevent ...
55
56	- a symbolically formed event like 'pmu/config=M,config1=N,config2=K/'
57	  where M, N, K are numbers (in decimal, hex, octal format).
58	  Acceptable values for each of 'config', 'config1' and 'config2'
59	  parameters are defined by corresponding entries in
60	  /sys/bus/event_source/devices/<pmu>/format/*
61
62	Note that the last two syntaxes support prefix and glob matching in
63	the PMU name to simplify creation of events across multiple instances
64	of the same type of PMU in large systems (e.g. memory controller PMUs).
65	Multiple PMU instances are typical for uncore PMUs, so the prefix
66	'uncore_' is also ignored when performing this match.
67
68
69-i::
70--no-inherit::
71        child tasks do not inherit counters
72-p::
73--pid=<pid>::
74        stat events on existing process id (comma separated list)
75
76-t::
77--tid=<tid>::
78        stat events on existing thread id (comma separated list)
79
80-b::
81--bpf-prog::
82        stat events on existing bpf program id (comma separated list),
83        requiring root rights. bpftool-prog could be used to find program
84        id all bpf programs in the system. For example:
85
86  # bpftool prog | head -n 1
87  17247: tracepoint  name sys_enter  tag 192d548b9d754067  gpl
88
89  # perf stat -e cycles,instructions --bpf-prog 17247 --timeout 1000
90
91   Performance counter stats for 'BPF program(s) 17247':
92
93             85,967      cycles
94             28,982      instructions              #    0.34  insn per cycle
95
96        1.102235068 seconds time elapsed
97
98--bpf-counters::
99	Use BPF programs to aggregate readings from perf_events.  This
100	allows multiple perf-stat sessions that are counting the same metric (cycles,
101	instructions, etc.) to share hardware counters.
102	To use BPF programs on common events by default, use
103	"perf config stat.bpf-counter-events=<list_of_events>".
104
105--bpf-attr-map::
106	With option "--bpf-counters", different perf-stat sessions share
107	information about shared BPF programs and maps via a pinned hashmap.
108	Use "--bpf-attr-map" to specify the path of this pinned hashmap.
109	The default path is /sys/fs/bpf/perf_attr_map.
110
111ifdef::HAVE_LIBPFM[]
112--pfm-events events::
113Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
114including support for event filters. For example '--pfm-events
115inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
116option using the comma separator. Hardware events and generic hardware
117events cannot be mixed together. The latter must be used with the -e
118option. The -e option and this one can be mixed and matched.  Events
119can be grouped using the {} notation.
120endif::HAVE_LIBPFM[]
121
122-a::
123--all-cpus::
124        system-wide collection from all CPUs (default if no target is specified)
125
126--no-scale::
127	Don't scale/normalize counter values
128
129-d::
130--detailed::
131	print more detailed statistics, can be specified up to 3 times
132
133	   -d:          detailed events, L1 and LLC data cache
134        -d -d:     more detailed events, dTLB and iTLB events
135     -d -d -d:     very detailed events, adding prefetch events
136
137-r::
138--repeat=<n>::
139	repeat command and print average + stddev (max: 100). 0 means forever.
140
141-B::
142--big-num::
143        print large numbers with thousands' separators according to locale.
144	Enabled by default. Use "--no-big-num" to disable.
145	Default setting can be changed with "perf config stat.big-num=false".
146
147-C::
148--cpu=::
149Count only on the list of CPUs provided. Multiple CPUs can be provided as a
150comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
151In per-thread mode, this option is ignored. The -a option is still necessary
152to activate system-wide monitoring. Default is to count on all CPUs.
153
154-A::
155--no-aggr::
156Do not aggregate counts across all monitored CPUs.
157
158-n::
159--null::
160null run - Don't start any counters.
161
162This can be useful to measure just elapsed wall-clock time - or to assess the
163raw overhead of perf stat itself, without running any counters.
164
165-v::
166--verbose::
167        be more verbose (show counter open errors, etc)
168
169-x SEP::
170--field-separator SEP::
171print counts using a CSV-style output to make it easy to import directly into
172spreadsheets. Columns are separated by the string specified in SEP.
173
174--table:: Display time for each run (-r option), in a table format, e.g.:
175
176  $ perf stat --null -r 5 --table perf bench sched pipe
177
178   Performance counter stats for 'perf bench sched pipe' (5 runs):
179
180             # Table of individual measurements:
181             5.189 (-0.293) #
182             5.189 (-0.294) #
183             5.186 (-0.296) #
184             5.663 (+0.181) ##
185             6.186 (+0.703) ####
186
187             # Final result:
188             5.483 +- 0.198 seconds time elapsed  ( +-  3.62% )
189
190-G name::
191--cgroup name::
192monitor only in the container (cgroup) called "name". This option is available only
193in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
194container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
195can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
196to first event, second cgroup to second event and so on. It is possible to provide
197an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
198corresponding events, i.e., they always refer to events defined earlier on the command
199line. If the user wants to track multiple events for a specific cgroup, the user can
200use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
201
202If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
203command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
204
205--for-each-cgroup name::
206Expand event list for each cgroup in "name" (allow multiple cgroups separated
207by comma).  It also support regex patterns to match multiple groups.  This has same
208effect that repeating -e option and -G option for each event x name.  This option
209cannot be used with -G/--cgroup option.
210
211-o file::
212--output file::
213Print the output into the designated file.
214
215--append::
216Append to the output file designated with the -o option. Ignored if -o is not specified.
217
218--log-fd::
219
220Log output to fd, instead of stderr.  Complementary to --output, and mutually exclusive
221with it.  --append may be used here.  Examples:
222     3>results  perf stat --log-fd 3          \-- $cmd
223     3>>results perf stat --log-fd 3 --append \-- $cmd
224
225--control=fifo:ctl-fifo[,ack-fifo]::
226--control=fd:ctl-fd[,ack-fd]::
227ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as follows.
228Listen on ctl-fd descriptor for command to control measurement ('enable': enable events,
229'disable': disable events). Measurements can be started with events disabled using
230--delay=-1 option. Optionally send control command completion ('ack\n') to ack-fd descriptor
231to synchronize with the controlling process. Example of bash shell script to enable and
232disable events during measurements:
233
234 #!/bin/bash
235
236 ctl_dir=/tmp/
237
238 ctl_fifo=${ctl_dir}perf_ctl.fifo
239 test -p ${ctl_fifo} && unlink ${ctl_fifo}
240 mkfifo ${ctl_fifo}
241 exec {ctl_fd}<>${ctl_fifo}
242
243 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
244 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
245 mkfifo ${ctl_ack_fifo}
246 exec {ctl_fd_ack}<>${ctl_ack_fifo}
247
248 perf stat -D -1 -e cpu-cycles -a -I 1000       \
249           --control fd:${ctl_fd},${ctl_fd_ack} \
250           \-- sleep 30 &
251 perf_pid=$!
252
253 sleep 5  && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
254 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
255
256 exec {ctl_fd_ack}>&-
257 unlink ${ctl_ack_fifo}
258
259 exec {ctl_fd}>&-
260 unlink ${ctl_fifo}
261
262 wait -n ${perf_pid}
263 exit $?
264
265
266--pre::
267--post::
268	Pre and post measurement hooks, e.g.:
269
270perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' \-- make -s -j64 O=defconfig-build/ bzImage
271
272-I msecs::
273--interval-print msecs::
274Print count deltas every N milliseconds (minimum: 1ms)
275The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals.  Use with caution.
276	example: 'perf stat -I 1000 -e cycles -a sleep 5'
277
278If the metric exists, it is calculated by the counts generated in this interval and the metric is printed after #.
279
280--interval-count times::
281Print count deltas for fixed number of times.
282This option should be used together with "-I" option.
283	example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
284
285--interval-clear::
286Clear the screen before next interval.
287
288--timeout msecs::
289Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
290This option is not supported with the "-I" option.
291	example: 'perf stat --time 2000 -e cycles -a'
292
293--metric-only::
294Only print computed metrics. Print them in a single line.
295Don't show any raw values. Not supported with --per-thread.
296
297--per-socket::
298Aggregate counts per processor socket for system-wide mode measurements.  This
299is a useful mode to detect imbalance between sockets.  To enable this mode,
300use --per-socket in addition to -a. (system-wide).  The output includes the
301socket number and the number of online processors on that socket. This is
302useful to gauge the amount of aggregation.
303
304--per-die::
305Aggregate counts per processor die for system-wide mode measurements.  This
306is a useful mode to detect imbalance between dies.  To enable this mode,
307use --per-die in addition to -a. (system-wide).  The output includes the
308die number and the number of online processors on that die. This is
309useful to gauge the amount of aggregation.
310
311--per-cluster::
312Aggregate counts per processor cluster for system-wide mode measurement.  This
313is a useful mode to detect imbalance between clusters.  To enable this mode,
314use --per-cluster in addition to -a. (system-wide).  The output includes the
315cluster number and the number of online processors on that cluster. This is
316useful to gauge the amount of aggregation. The information of cluster ID and
317related CPUs can be gotten from /sys/devices/system/cpu/cpuX/topology/cluster_{id, cpus}.
318
319--per-cache::
320Aggregate counts per cache instance for system-wide mode measurements.  By
321default, the aggregation happens for the cache level at the highest index
322in the system. To specify a particular level, mention the cache level
323alongside the option in the format [Ll][1-9][0-9]*. For example:
324Using option "--per-cache=l3" or "--per-cache=L3" will aggregate the
325information at the boundary of the level 3 cache in the system.
326
327--per-core::
328Aggregate counts per physical processor for system-wide mode measurements.  This
329is a useful mode to detect imbalance between physical cores.  To enable this mode,
330use --per-core in addition to -a. (system-wide).  The output includes the
331core number and the number of online logical processors on that physical processor.
332
333--per-thread::
334Aggregate counts per monitored threads, when monitoring threads (-t option)
335or processes (-p option).
336
337--per-node::
338Aggregate counts per NUMA nodes for system-wide mode measurements. This
339is a useful mode to detect imbalance between NUMA nodes. To enable this
340mode, use --per-node in addition to -a. (system-wide).
341
342-D msecs::
343--delay msecs::
344After starting the program, wait msecs before measuring (-1: start with events
345disabled). This is useful to filter out the startup phase of the program,
346which is often very different.
347
348-T::
349--transaction::
350
351Print statistics of transactional execution if supported.
352
353--metric-no-group::
354By default, events to compute a metric are placed in weak groups. The
355group tries to enforce scheduling all or none of the events. The
356--metric-no-group option places events outside of groups and may
357increase the chance of the event being scheduled - leading to more
358accuracy. However, as events may not be scheduled together accuracy
359for metrics like instructions per cycle can be lower - as both metrics
360may no longer be being measured at the same time.
361
362--metric-no-merge::
363By default metric events in different weak groups can be shared if one
364group contains all the events needed by another. In such cases one
365group will be eliminated reducing event multiplexing and making it so
366that certain groups of metrics sum to 100%. A downside to sharing a
367group is that the group may require multiplexing and so accuracy for a
368small group that need not have multiplexing is lowered. This option
369forbids the event merging logic from sharing events between groups and
370may be used to increase accuracy in this case.
371
372--metric-no-threshold::
373Metric thresholds may increase the number of events necessary to
374compute whether a metric has exceeded its threshold expression. This
375may not be desirable, for example, as the events can introduce
376multiplexing. This option disables the adding of threshold expression
377events for a metric. However, if there are sufficient events to
378compute the threshold then the threshold is still computed and used to
379color the metric's computed value.
380
381--quiet::
382Don't print output, warnings or messages. This is useful with perf stat
383record below to only write data to the perf.data file.
384
385STAT RECORD
386-----------
387Stores stat data into perf data file.
388
389-o file::
390--output file::
391Output file name.
392
393STAT REPORT
394-----------
395Reads and reports stat data from perf data file.
396
397-i file::
398--input file::
399Input file name.
400
401--per-socket::
402Aggregate counts per processor socket for system-wide mode measurements.
403
404--per-die::
405Aggregate counts per processor die for system-wide mode measurements.
406
407--per-cluster::
408Aggregate counts perf processor cluster for system-wide mode measurements.
409
410--per-cache::
411Aggregate counts per cache instance for system-wide mode measurements.  By
412default, the aggregation happens for the cache level at the highest index
413in the system. To specify a particular level, mention the cache level
414alongside the option in the format [Ll][1-9][0-9]*. For example: Using
415option "--per-cache=l3" or "--per-cache=L3" will aggregate the
416information at the boundary of the level 3 cache in the system.
417
418--per-core::
419Aggregate counts per physical processor for system-wide mode measurements.
420
421-M::
422--metrics::
423Print metrics or metricgroups specified in a comma separated list.
424For a group all metrics from the group are added.
425The events from the metrics are automatically measured.
426See perf list output for the possible metrics and metricgroups.
427
428	When threshold information is available for a metric, the
429	color red is used to signify a metric has exceeded a threshold
430	while green shows it hasn't. The default color means that
431	no threshold information was available or the threshold
432	couldn't be computed.
433
434-A::
435--no-aggr::
436--no-merge::
437Do not aggregate/merge counts across monitored CPUs or PMUs.
438
439When multiple events are created from a single event specification,
440stat will, by default, aggregate the event counts and show the result
441in a single row. This option disables that behavior and shows the
442individual events and counts.
443
444Multiple events are created from a single event specification when:
445
4461. PID monitoring isn't requested and the system has more than one
447   CPU. For example, a system with 8 SMT threads will have one event
448   opened on each thread and aggregation is performed across them.
449
4502. Prefix or glob wildcard matching is used for the PMU name. For
451   example, multiple memory controller PMUs may exist typically with a
452   suffix of _0, _1, etc. By default the event counts will all be
453   combined if the PMU is specified without the suffix such as
454   uncore_imc rather than uncore_imc_0.
455
4563. Aliases, which are listed immediately after the Kernel PMU events
457   by perf list, are used.
458
459--hybrid-merge::
460Merge core event counts from all core PMUs. In hybrid or big.LITTLE
461systems by default each core PMU will report its count
462separately. This option forces core PMU counts to be combined to give
463a behavior closer to having a single CPU type in the system.
464
465--topdown::
466Print top-down metrics supported by the CPU. This allows to determine
467bottle necks in the CPU pipeline for CPU bound workloads, by breaking
468the cycles consumed down into frontend bound, backend bound, bad
469speculation and retiring.
470
471Frontend bound means that the CPU cannot fetch and decode instructions fast
472enough. Backend bound means that computation or memory access is the bottle
473neck. Bad Speculation means that the CPU wasted cycles due to branch
474mispredictions and similar issues. Retiring means that the CPU computed without
475an apparently bottleneck. The bottleneck is only the real bottleneck
476if the workload is actually bound by the CPU and not by something else.
477
478For best results it is usually a good idea to use it with interval
479mode like -I 1000, as the bottleneck of workloads can change often.
480
481This enables --metric-only, unless overridden with --no-metric-only.
482
483The following restrictions only apply to older Intel CPUs and Atom,
484on newer CPUs (IceLake and later) TopDown can be collected for any thread:
485
486The top down metrics are collected per core instead of per
487CPU thread. Per core mode is automatically enabled
488and -a (global monitoring) is needed, requiring root rights or
489perf.perf_event_paranoid=-1.
490
491Topdown uses the full Performance Monitoring Unit, and needs
492disabling of the NMI watchdog (as root):
493echo 0 > /proc/sys/kernel/nmi_watchdog
494for best results. Otherwise the bottlenecks may be inconsistent
495on workload with changing phases.
496
497To interpret the results it is usually needed to know on which
498CPUs the workload runs on. If needed the CPUs can be forced using
499taskset.
500
501--record-tpebs::
502Enable automatic sampling on Intel TPEBS retire_latency events (event with :R
503modifier). Without this option, perf would not capture dynamic retire_latency
504at runtime. Currently, a zero value is assigned to the retire_latency event when
505this option is not set. The TPEBS hardware feature starts from Intel Granite
506Rapids microarchitecture. This option only exists in X86_64 and is meaningful on
507Intel platforms with TPEBS feature.
508
509--td-level::
510Print the top-down statistics that equal the input level. It allows
511users to print the interested top-down metrics level instead of the
512level 1 top-down metrics.
513
514As the higher levels gather more metrics and use more counters they
515will be less accurate. By convention a metric can be examined by
516appending '_group' to it and this will increase accuracy compared to
517gathering all metrics for a level. For example, level 1 analysis may
518highlight 'tma_frontend_bound'. This metric may be drilled into with
519'tma_frontend_bound_group' with
520'perf stat -M tma_frontend_bound_group...'.
521
522Error out if the input is higher than the supported max level.
523
524--smi-cost::
525Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
526
527During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
528freeze core counters on SMI.
529The aperf counter will not be effected by the setting.
530The cost of SMI can be measured by (aperf - unhalted core cycles).
531
532In practice, the percentages of SMI cycles is very useful for performance
533oriented analysis. --metric_only will be applied by default.
534The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
535
536Users who wants to get the actual value can apply --no-metric-only.
537
538--all-kernel::
539Configure all used events to run in kernel space.
540
541--all-user::
542Configure all used events to run in user space.
543
544--percore-show-thread::
545The event modifier "percore" has supported to sum up the event counts
546for all hardware threads in a core and show the counts per core.
547
548This option with event modifier "percore" enabled also sums up the event
549counts for all hardware threads in a core but show the sum counts per
550hardware thread. This is essentially a replacement for the any bit and
551convenient for post processing.
552
553--summary::
554Print summary for interval mode (-I).
555
556--no-csv-summary::
557Don't print 'summary' at the first column for CVS summary output.
558This option must be used with -x and --summary.
559
560This option can be enabled in perf config by setting the variable
561'stat.no-csv-summary'.
562
563$ perf config stat.no-csv-summary=true
564
565--cputype::
566Only enable events on applying cpu with this type for hybrid platform
567(e.g. core or atom)"
568
569EXAMPLES
570--------
571
572$ perf stat \-- make
573
574   Performance counter stats for 'make':
575
576        83723.452481      task-clock:u (msec)       #    1.004 CPUs utilized
577                   0      context-switches:u        #    0.000 K/sec
578                   0      cpu-migrations:u          #    0.000 K/sec
579           3,228,188      page-faults:u             #    0.039 M/sec
580     229,570,665,834      cycles:u                  #    2.742 GHz
581     313,163,853,778      instructions:u            #    1.36  insn per cycle
582      69,704,684,856      branches:u                #  832.559 M/sec
583       2,078,861,393      branch-misses:u           #    2.98% of all branches
584
585        83.409183620 seconds time elapsed
586
587        74.684747000 seconds user
588         8.739217000 seconds sys
589
590TIMINGS
591-------
592As displayed in the example above we can display 3 types of timings.
593We always display the time the counters were enabled/alive:
594
595        83.409183620 seconds time elapsed
596
597For workload sessions we also display time the workloads spent in
598user/system lands:
599
600        74.684747000 seconds user
601         8.739217000 seconds sys
602
603Those times are the very same as displayed by the 'time' tool.
604
605CSV FORMAT
606----------
607
608With -x, perf stat is able to output a not-quite-CSV format output
609Commas in the output are not put into "". To make it easy to parse
610it is recommended to use a different character like -x \;
611
612The fields are in this order:
613
614	- optional usec time stamp in fractions of second (with -I xxx)
615	- optional CPU, core, or socket identifier
616	- optional number of logical CPUs aggregated
617	- counter value
618	- unit of the counter value or empty
619	- event name
620	- run time of counter
621	- percentage of measurement time the counter was running
622	- optional variance if multiple values are collected with -r
623	- optional metric value
624	- optional unit of metric
625
626Additional metrics may be printed with all earlier fields being empty.
627
628include::intel-hybrid.txt[]
629
630JSON FORMAT
631-----------
632
633With -j, perf stat is able to print out a JSON format output
634that can be used for parsing.
635
636- timestamp : optional usec time stamp in fractions of second (with -I)
637- optional aggregate options:
638		- core : core identifier (with --per-core)
639		- die : die identifier (with --per-die)
640		- socket : socket identifier (with --per-socket)
641		- node : node identifier (with --per-node)
642		- thread : thread identifier (with --per-thread)
643- counter-value : counter value
644- unit : unit of the counter value or empty
645- event : event name
646- variance : optional variance if multiple values are collected (with -r)
647- runtime : run time of counter
648- metric-value : optional metric value
649- metric-unit : optional unit of metric
650
651SEE ALSO
652--------
653linkperf:perf-top[1], linkperf:perf-list[1]
654