Loading...
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 (eventsel+umask) in the form of rNNN where NNN is a
40 hexadecimal event descriptor.
41
42 - a symbolically formed event like 'pmu/param1=0x3,param2/' where
43 param1 and param2 are defined as formats for the PMU in
44 /sys/bus/event_source/devices/<pmu>/format/*
45
46 - a symbolically formed event like 'pmu/config=M,config1=N,config2=K/'
47 where M, N, K are numbers (in decimal, hex, octal format).
48 Acceptable values for each of 'config', 'config1' and 'config2'
49 parameters are defined by corresponding entries in
50 /sys/bus/event_source/devices/<pmu>/format/*
51
52 Note that the last two syntaxes support prefix and glob matching in
53 the PMU name to simplify creation of events accross multiple instances
54 of the same type of PMU in large systems (e.g. memory controller PMUs).
55 Multiple PMU instances are typical for uncore PMUs, so the prefix
56 'uncore_' is also ignored when performing this match.
57
58
59-i::
60--no-inherit::
61 child tasks do not inherit counters
62-p::
63--pid=<pid>::
64 stat events on existing process id (comma separated list)
65
66-t::
67--tid=<tid>::
68 stat events on existing thread id (comma separated list)
69
70
71-a::
72--all-cpus::
73 system-wide collection from all CPUs (default if no target is specified)
74
75-c::
76--scale::
77 scale/normalize counter values
78
79-d::
80--detailed::
81 print more detailed statistics, can be specified up to 3 times
82
83 -d: detailed events, L1 and LLC data cache
84 -d -d: more detailed events, dTLB and iTLB events
85 -d -d -d: very detailed events, adding prefetch events
86
87-r::
88--repeat=<n>::
89 repeat command and print average + stddev (max: 100). 0 means forever.
90
91-B::
92--big-num::
93 print large numbers with thousands' separators according to locale
94
95-C::
96--cpu=::
97Count only on the list of CPUs provided. Multiple CPUs can be provided as a
98comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
99In per-thread mode, this option is ignored. The -a option is still necessary
100to activate system-wide monitoring. Default is to count on all CPUs.
101
102-A::
103--no-aggr::
104Do not aggregate counts across all monitored CPUs.
105
106-n::
107--null::
108 null run - don't start any counters
109
110-v::
111--verbose::
112 be more verbose (show counter open errors, etc)
113
114-x SEP::
115--field-separator SEP::
116print counts using a CSV-style output to make it easy to import directly into
117spreadsheets. Columns are separated by the string specified in SEP.
118
119-G name::
120--cgroup name::
121monitor only in the container (cgroup) called "name". This option is available only
122in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
123container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
124can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
125to first event, second cgroup to second event and so on. It is possible to provide
126an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
127corresponding events, i.e., they always refer to events defined earlier on the command
128line. If the user wants to track multiple events for a specific cgroup, the user can
129use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
130
131If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
132command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
133
134-o file::
135--output file::
136Print the output into the designated file.
137
138--append::
139Append to the output file designated with the -o option. Ignored if -o is not specified.
140
141--log-fd::
142
143Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
144with it. --append may be used here. Examples:
145 3>results perf stat --log-fd 3 -- $cmd
146 3>>results perf stat --log-fd 3 --append -- $cmd
147
148--pre::
149--post::
150 Pre and post measurement hooks, e.g.:
151
152perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
153
154-I msecs::
155--interval-print msecs::
156Print count deltas every N milliseconds (minimum: 1ms)
157The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
158 example: 'perf stat -I 1000 -e cycles -a sleep 5'
159
160--interval-count times::
161Print count deltas for fixed number of times.
162This option should be used together with "-I" option.
163 example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
164
165--timeout msecs::
166Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
167This option is not supported with the "-I" option.
168 example: 'perf stat --time 2000 -e cycles -a'
169
170--metric-only::
171Only print computed metrics. Print them in a single line.
172Don't show any raw values. Not supported with --per-thread.
173
174--per-socket::
175Aggregate counts per processor socket for system-wide mode measurements. This
176is a useful mode to detect imbalance between sockets. To enable this mode,
177use --per-socket in addition to -a. (system-wide). The output includes the
178socket number and the number of online processors on that socket. This is
179useful to gauge the amount of aggregation.
180
181--per-core::
182Aggregate counts per physical processor for system-wide mode measurements. This
183is a useful mode to detect imbalance between physical cores. To enable this mode,
184use --per-core in addition to -a. (system-wide). The output includes the
185core number and the number of online logical processors on that physical processor.
186
187--per-thread::
188Aggregate counts per monitored threads, when monitoring threads (-t option)
189or processes (-p option).
190
191-D msecs::
192--delay msecs::
193After starting the program, wait msecs before measuring. This is useful to
194filter out the startup phase of the program, which is often very different.
195
196-T::
197--transaction::
198
199Print statistics of transactional execution if supported.
200
201STAT RECORD
202-----------
203Stores stat data into perf data file.
204
205-o file::
206--output file::
207Output file name.
208
209STAT REPORT
210-----------
211Reads and reports stat data from perf data file.
212
213-i file::
214--input file::
215Input file name.
216
217--per-socket::
218Aggregate counts per processor socket for system-wide mode measurements.
219
220--per-core::
221Aggregate counts per physical processor for system-wide mode measurements.
222
223-M::
224--metrics::
225Print metrics or metricgroups specified in a comma separated list.
226For a group all metrics from the group are added.
227The events from the metrics are automatically measured.
228See perf list output for the possble metrics and metricgroups.
229
230-A::
231--no-aggr::
232Do not aggregate counts across all monitored CPUs.
233
234--topdown::
235Print top down level 1 metrics if supported by the CPU. This allows to
236determine bottle necks in the CPU pipeline for CPU bound workloads,
237by breaking the cycles consumed down into frontend bound, backend bound,
238bad speculation and retiring.
239
240Frontend bound means that the CPU cannot fetch and decode instructions fast
241enough. Backend bound means that computation or memory access is the bottle
242neck. Bad Speculation means that the CPU wasted cycles due to branch
243mispredictions and similar issues. Retiring means that the CPU computed without
244an apparently bottleneck. The bottleneck is only the real bottleneck
245if the workload is actually bound by the CPU and not by something else.
246
247For best results it is usually a good idea to use it with interval
248mode like -I 1000, as the bottleneck of workloads can change often.
249
250The top down metrics are collected per core instead of per
251CPU thread. Per core mode is automatically enabled
252and -a (global monitoring) is needed, requiring root rights or
253perf.perf_event_paranoid=-1.
254
255Topdown uses the full Performance Monitoring Unit, and needs
256disabling of the NMI watchdog (as root):
257echo 0 > /proc/sys/kernel/nmi_watchdog
258for best results. Otherwise the bottlenecks may be inconsistent
259on workload with changing phases.
260
261This enables --metric-only, unless overriden with --no-metric-only.
262
263To interpret the results it is usually needed to know on which
264CPUs the workload runs on. If needed the CPUs can be forced using
265taskset.
266
267--no-merge::
268Do not merge results from same PMUs.
269
270When multiple events are created from a single event specification,
271stat will, by default, aggregate the event counts and show the result
272in a single row. This option disables that behavior and shows
273the individual events and counts.
274
275Multiple events are created from a single event specification when:
2761. Prefix or glob matching is used for the PMU name.
2772. Aliases, which are listed immediately after the Kernel PMU events
278 by perf list, are used.
279
280--smi-cost::
281Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
282
283During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
284freeze core counters on SMI.
285The aperf counter will not be effected by the setting.
286The cost of SMI can be measured by (aperf - unhalted core cycles).
287
288In practice, the percentages of SMI cycles is very useful for performance
289oriented analysis. --metric_only will be applied by default.
290The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
291
292Users who wants to get the actual value can apply --no-metric-only.
293
294EXAMPLES
295--------
296
297$ perf stat -- make -j
298
299 Performance counter stats for 'make -j':
300
301 8117.370256 task clock ticks # 11.281 CPU utilization factor
302 678 context switches # 0.000 M/sec
303 133 CPU migrations # 0.000 M/sec
304 235724 pagefaults # 0.029 M/sec
305 24821162526 CPU cycles # 3057.784 M/sec
306 18687303457 instructions # 2302.138 M/sec
307 172158895 cache references # 21.209 M/sec
308 27075259 cache misses # 3.335 M/sec
309
310 Wall-clock time elapsed: 719.554352 msecs
311
312CSV FORMAT
313----------
314
315With -x, perf stat is able to output a not-quite-CSV format output
316Commas in the output are not put into "". To make it easy to parse
317it is recommended to use a different character like -x \;
318
319The fields are in this order:
320
321 - optional usec time stamp in fractions of second (with -I xxx)
322 - optional CPU, core, or socket identifier
323 - optional number of logical CPUs aggregated
324 - counter value
325 - unit of the counter value or empty
326 - event name
327 - run time of counter
328 - percentage of measurement time the counter was running
329 - optional variance if multiple values are collected with -r
330 - optional metric value
331 - optional unit of metric
332
333Additional metrics may be printed with all earlier fields being empty.
334
335SEE ALSO
336--------
337linkperf:perf-top[1], linkperf:perf-list[1]
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 (eventsel+umask) in the form of rNNN where NNN is a
40 hexadecimal event descriptor.
41
42 - a symbolic or raw PMU event followed by an optional colon
43 and a list of event modifiers, e.g., cpu-cycles:p. See the
44 linkperf:perf-list[1] man page for details on event modifiers.
45
46 - a symbolically formed event like 'pmu/param1=0x3,param2/' where
47 param1 and param2 are defined as formats for the PMU in
48 /sys/bus/event_source/devices/<pmu>/format/*
49
50 'percore' is a event qualifier that sums up the event counts for both
51 hardware threads in a core. For example:
52 perf stat -A -a -e cpu/event,percore=1/,otherevent ...
53
54 - a symbolically formed event like 'pmu/config=M,config1=N,config2=K/'
55 where M, N, K are numbers (in decimal, hex, octal format).
56 Acceptable values for each of 'config', 'config1' and 'config2'
57 parameters are defined by corresponding entries in
58 /sys/bus/event_source/devices/<pmu>/format/*
59
60 Note that the last two syntaxes support prefix and glob matching in
61 the PMU name to simplify creation of events across multiple instances
62 of the same type of PMU in large systems (e.g. memory controller PMUs).
63 Multiple PMU instances are typical for uncore PMUs, so the prefix
64 'uncore_' is also ignored when performing this match.
65
66
67-i::
68--no-inherit::
69 child tasks do not inherit counters
70-p::
71--pid=<pid>::
72 stat events on existing process id (comma separated list)
73
74-t::
75--tid=<tid>::
76 stat events on existing thread id (comma separated list)
77
78-b::
79--bpf-prog::
80 stat events on existing bpf program id (comma separated list),
81 requiring root rights. bpftool-prog could be used to find program
82 id all bpf programs in the system. For example:
83
84 # bpftool prog | head -n 1
85 17247: tracepoint name sys_enter tag 192d548b9d754067 gpl
86
87 # perf stat -e cycles,instructions --bpf-prog 17247 --timeout 1000
88
89 Performance counter stats for 'BPF program(s) 17247':
90
91 85,967 cycles
92 28,982 instructions # 0.34 insn per cycle
93
94 1.102235068 seconds time elapsed
95
96--bpf-counters::
97 Use BPF programs to aggregate readings from perf_events. This
98 allows multiple perf-stat sessions that are counting the same metric (cycles,
99 instructions, etc.) to share hardware counters.
100 To use BPF programs on common events by default, use
101 "perf config stat.bpf-counter-events=<list_of_events>".
102
103--bpf-attr-map::
104 With option "--bpf-counters", different perf-stat sessions share
105 information about shared BPF programs and maps via a pinned hashmap.
106 Use "--bpf-attr-map" to specify the path of this pinned hashmap.
107 The default path is /sys/fs/bpf/perf_attr_map.
108
109ifdef::HAVE_LIBPFM[]
110--pfm-events events::
111Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
112including support for event filters. For example '--pfm-events
113inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
114option using the comma separator. Hardware events and generic hardware
115events cannot be mixed together. The latter must be used with the -e
116option. The -e option and this one can be mixed and matched. Events
117can be grouped using the {} notation.
118endif::HAVE_LIBPFM[]
119
120-a::
121--all-cpus::
122 system-wide collection from all CPUs (default if no target is specified)
123
124--no-scale::
125 Don't scale/normalize counter values
126
127-d::
128--detailed::
129 print more detailed statistics, can be specified up to 3 times
130
131 -d: detailed events, L1 and LLC data cache
132 -d -d: more detailed events, dTLB and iTLB events
133 -d -d -d: very detailed events, adding prefetch events
134
135-r::
136--repeat=<n>::
137 repeat command and print average + stddev (max: 100). 0 means forever.
138
139-B::
140--big-num::
141 print large numbers with thousands' separators according to locale.
142 Enabled by default. Use "--no-big-num" to disable.
143 Default setting can be changed with "perf config stat.big-num=false".
144
145-C::
146--cpu=::
147Count only on the list of CPUs provided. Multiple CPUs can be provided as a
148comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
149In per-thread mode, this option is ignored. The -a option is still necessary
150to activate system-wide monitoring. Default is to count on all CPUs.
151
152-A::
153--no-aggr::
154Do not aggregate counts across all monitored CPUs.
155
156-n::
157--null::
158null run - Don't start any counters.
159
160This can be useful to measure just elapsed wall-clock time - or to assess the
161raw overhead of perf stat itself, without running any counters.
162
163-v::
164--verbose::
165 be more verbose (show counter open errors, etc)
166
167-x SEP::
168--field-separator SEP::
169print counts using a CSV-style output to make it easy to import directly into
170spreadsheets. Columns are separated by the string specified in SEP.
171
172--table:: Display time for each run (-r option), in a table format, e.g.:
173
174 $ perf stat --null -r 5 --table perf bench sched pipe
175
176 Performance counter stats for 'perf bench sched pipe' (5 runs):
177
178 # Table of individual measurements:
179 5.189 (-0.293) #
180 5.189 (-0.294) #
181 5.186 (-0.296) #
182 5.663 (+0.181) ##
183 6.186 (+0.703) ####
184
185 # Final result:
186 5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
187
188-G name::
189--cgroup name::
190monitor only in the container (cgroup) called "name". This option is available only
191in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
192container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
193can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
194to first event, second cgroup to second event and so on. It is possible to provide
195an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
196corresponding events, i.e., they always refer to events defined earlier on the command
197line. If the user wants to track multiple events for a specific cgroup, the user can
198use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
199
200If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
201command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
202
203--for-each-cgroup name::
204Expand event list for each cgroup in "name" (allow multiple cgroups separated
205by comma). It also support regex patterns to match multiple groups. This has same
206effect that repeating -e option and -G option for each event x name. This option
207cannot be used with -G/--cgroup option.
208
209-o file::
210--output file::
211Print the output into the designated file.
212
213--append::
214Append to the output file designated with the -o option. Ignored if -o is not specified.
215
216--log-fd::
217
218Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
219with it. --append may be used here. Examples:
220 3>results perf stat --log-fd 3 -- $cmd
221 3>>results perf stat --log-fd 3 --append -- $cmd
222
223--control=fifo:ctl-fifo[,ack-fifo]::
224--control=fd:ctl-fd[,ack-fd]::
225ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as follows.
226Listen on ctl-fd descriptor for command to control measurement ('enable': enable events,
227'disable': disable events). Measurements can be started with events disabled using
228--delay=-1 option. Optionally send control command completion ('ack\n') to ack-fd descriptor
229to synchronize with the controlling process. Example of bash shell script to enable and
230disable events during measurements:
231
232 #!/bin/bash
233
234 ctl_dir=/tmp/
235
236 ctl_fifo=${ctl_dir}perf_ctl.fifo
237 test -p ${ctl_fifo} && unlink ${ctl_fifo}
238 mkfifo ${ctl_fifo}
239 exec {ctl_fd}<>${ctl_fifo}
240
241 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
242 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
243 mkfifo ${ctl_ack_fifo}
244 exec {ctl_fd_ack}<>${ctl_ack_fifo}
245
246 perf stat -D -1 -e cpu-cycles -a -I 1000 \
247 --control fd:${ctl_fd},${ctl_fd_ack} \
248 -- sleep 30 &
249 perf_pid=$!
250
251 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
252 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
253
254 exec {ctl_fd_ack}>&-
255 unlink ${ctl_ack_fifo}
256
257 exec {ctl_fd}>&-
258 unlink ${ctl_fifo}
259
260 wait -n ${perf_pid}
261 exit $?
262
263
264--pre::
265--post::
266 Pre and post measurement hooks, e.g.:
267
268perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
269
270-I msecs::
271--interval-print msecs::
272Print count deltas every N milliseconds (minimum: 1ms)
273The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
274 example: 'perf stat -I 1000 -e cycles -a sleep 5'
275
276If the metric exists, it is calculated by the counts generated in this interval and the metric is printed after #.
277
278--interval-count times::
279Print count deltas for fixed number of times.
280This option should be used together with "-I" option.
281 example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
282
283--interval-clear::
284Clear the screen before next interval.
285
286--timeout msecs::
287Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
288This option is not supported with the "-I" option.
289 example: 'perf stat --time 2000 -e cycles -a'
290
291--metric-only::
292Only print computed metrics. Print them in a single line.
293Don't show any raw values. Not supported with --per-thread.
294
295--per-socket::
296Aggregate counts per processor socket for system-wide mode measurements. This
297is a useful mode to detect imbalance between sockets. To enable this mode,
298use --per-socket in addition to -a. (system-wide). The output includes the
299socket number and the number of online processors on that socket. This is
300useful to gauge the amount of aggregation.
301
302--per-die::
303Aggregate counts per processor die for system-wide mode measurements. This
304is a useful mode to detect imbalance between dies. To enable this mode,
305use --per-die in addition to -a. (system-wide). The output includes the
306die number and the number of online processors on that die. This is
307useful to gauge the amount of aggregation.
308
309--per-core::
310Aggregate counts per physical processor for system-wide mode measurements. This
311is a useful mode to detect imbalance between physical cores. To enable this mode,
312use --per-core in addition to -a. (system-wide). The output includes the
313core number and the number of online logical processors on that physical processor.
314
315--per-thread::
316Aggregate counts per monitored threads, when monitoring threads (-t option)
317or processes (-p option).
318
319--per-node::
320Aggregate counts per NUMA nodes for system-wide mode measurements. This
321is a useful mode to detect imbalance between NUMA nodes. To enable this
322mode, use --per-node in addition to -a. (system-wide).
323
324-D msecs::
325--delay msecs::
326After starting the program, wait msecs before measuring (-1: start with events
327disabled). This is useful to filter out the startup phase of the program,
328which is often very different.
329
330-T::
331--transaction::
332
333Print statistics of transactional execution if supported.
334
335--metric-no-group::
336By default, events to compute a metric are placed in weak groups. The
337group tries to enforce scheduling all or none of the events. The
338--metric-no-group option places events outside of groups and may
339increase the chance of the event being scheduled - leading to more
340accuracy. However, as events may not be scheduled together accuracy
341for metrics like instructions per cycle can be lower - as both metrics
342may no longer be being measured at the same time.
343
344--metric-no-merge::
345By default metric events in different weak groups can be shared if one
346group contains all the events needed by another. In such cases one
347group will be eliminated reducing event multiplexing and making it so
348that certain groups of metrics sum to 100%. A downside to sharing a
349group is that the group may require multiplexing and so accuracy for a
350small group that need not have multiplexing is lowered. This option
351forbids the event merging logic from sharing events between groups and
352may be used to increase accuracy in this case.
353
354--quiet::
355Don't print output. This is useful with perf stat record below to only
356write data to the perf.data file.
357
358STAT RECORD
359-----------
360Stores stat data into perf data file.
361
362-o file::
363--output file::
364Output file name.
365
366STAT REPORT
367-----------
368Reads and reports stat data from perf data file.
369
370-i file::
371--input file::
372Input file name.
373
374--per-socket::
375Aggregate counts per processor socket for system-wide mode measurements.
376
377--per-die::
378Aggregate counts per processor die for system-wide mode measurements.
379
380--per-core::
381Aggregate counts per physical processor for system-wide mode measurements.
382
383-M::
384--metrics::
385Print metrics or metricgroups specified in a comma separated list.
386For a group all metrics from the group are added.
387The events from the metrics are automatically measured.
388See perf list output for the possble metrics and metricgroups.
389
390-A::
391--no-aggr::
392Do not aggregate counts across all monitored CPUs.
393
394--topdown::
395Print complete top-down metrics supported by the CPU. This allows to
396determine bottle necks in the CPU pipeline for CPU bound workloads,
397by breaking the cycles consumed down into frontend bound, backend bound,
398bad speculation and retiring.
399
400Frontend bound means that the CPU cannot fetch and decode instructions fast
401enough. Backend bound means that computation or memory access is the bottle
402neck. Bad Speculation means that the CPU wasted cycles due to branch
403mispredictions and similar issues. Retiring means that the CPU computed without
404an apparently bottleneck. The bottleneck is only the real bottleneck
405if the workload is actually bound by the CPU and not by something else.
406
407For best results it is usually a good idea to use it with interval
408mode like -I 1000, as the bottleneck of workloads can change often.
409
410This enables --metric-only, unless overridden with --no-metric-only.
411
412The following restrictions only apply to older Intel CPUs and Atom,
413on newer CPUs (IceLake and later) TopDown can be collected for any thread:
414
415The top down metrics are collected per core instead of per
416CPU thread. Per core mode is automatically enabled
417and -a (global monitoring) is needed, requiring root rights or
418perf.perf_event_paranoid=-1.
419
420Topdown uses the full Performance Monitoring Unit, and needs
421disabling of the NMI watchdog (as root):
422echo 0 > /proc/sys/kernel/nmi_watchdog
423for best results. Otherwise the bottlenecks may be inconsistent
424on workload with changing phases.
425
426To interpret the results it is usually needed to know on which
427CPUs the workload runs on. If needed the CPUs can be forced using
428taskset.
429
430--td-level::
431Print the top-down statistics that equal to or lower than the input level.
432It allows users to print the interested top-down metrics level instead of
433the complete top-down metrics.
434
435The availability of the top-down metrics level depends on the hardware. For
436example, Ice Lake only supports L1 top-down metrics. The Sapphire Rapids
437supports both L1 and L2 top-down metrics.
438
439Default: 0 means the max level that the current hardware support.
440Error out if the input is higher than the supported max level.
441
442--no-merge::
443Do not merge results from same PMUs.
444
445When multiple events are created from a single event specification,
446stat will, by default, aggregate the event counts and show the result
447in a single row. This option disables that behavior and shows
448the individual events and counts.
449
450Multiple events are created from a single event specification when:
4511. Prefix or glob matching is used for the PMU name.
4522. Aliases, which are listed immediately after the Kernel PMU events
453 by perf list, are used.
454
455--smi-cost::
456Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
457
458During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
459freeze core counters on SMI.
460The aperf counter will not be effected by the setting.
461The cost of SMI can be measured by (aperf - unhalted core cycles).
462
463In practice, the percentages of SMI cycles is very useful for performance
464oriented analysis. --metric_only will be applied by default.
465The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
466
467Users who wants to get the actual value can apply --no-metric-only.
468
469--all-kernel::
470Configure all used events to run in kernel space.
471
472--all-user::
473Configure all used events to run in user space.
474
475--percore-show-thread::
476The event modifier "percore" has supported to sum up the event counts
477for all hardware threads in a core and show the counts per core.
478
479This option with event modifier "percore" enabled also sums up the event
480counts for all hardware threads in a core but show the sum counts per
481hardware thread. This is essentially a replacement for the any bit and
482convenient for post processing.
483
484--summary::
485Print summary for interval mode (-I).
486
487--no-csv-summary::
488Don't print 'summary' at the first column for CVS summary output.
489This option must be used with -x and --summary.
490
491This option can be enabled in perf config by setting the variable
492'stat.no-csv-summary'.
493
494$ perf config stat.no-csv-summary=true
495
496EXAMPLES
497--------
498
499$ perf stat -- make
500
501 Performance counter stats for 'make':
502
503 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
504 0 context-switches:u # 0.000 K/sec
505 0 cpu-migrations:u # 0.000 K/sec
506 3,228,188 page-faults:u # 0.039 M/sec
507 229,570,665,834 cycles:u # 2.742 GHz
508 313,163,853,778 instructions:u # 1.36 insn per cycle
509 69,704,684,856 branches:u # 832.559 M/sec
510 2,078,861,393 branch-misses:u # 2.98% of all branches
511
512 83.409183620 seconds time elapsed
513
514 74.684747000 seconds user
515 8.739217000 seconds sys
516
517TIMINGS
518-------
519As displayed in the example above we can display 3 types of timings.
520We always display the time the counters were enabled/alive:
521
522 83.409183620 seconds time elapsed
523
524For workload sessions we also display time the workloads spent in
525user/system lands:
526
527 74.684747000 seconds user
528 8.739217000 seconds sys
529
530Those times are the very same as displayed by the 'time' tool.
531
532CSV FORMAT
533----------
534
535With -x, perf stat is able to output a not-quite-CSV format output
536Commas in the output are not put into "". To make it easy to parse
537it is recommended to use a different character like -x \;
538
539The fields are in this order:
540
541 - optional usec time stamp in fractions of second (with -I xxx)
542 - optional CPU, core, or socket identifier
543 - optional number of logical CPUs aggregated
544 - counter value
545 - unit of the counter value or empty
546 - event name
547 - run time of counter
548 - percentage of measurement time the counter was running
549 - optional variance if multiple values are collected with -r
550 - optional metric value
551 - optional unit of metric
552
553Additional metrics may be printed with all earlier fields being empty.
554
555include::intel-hybrid.txt[]
556
557SEE ALSO
558--------
559linkperf:perf-top[1], linkperf:perf-list[1]