xref: /linux/tools/perf/Documentation/perf-stat.txt (revision c53b4bb02b45ceec7a590e47820afbb5cef0bb81)
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-i::
53--no-inherit::
54        child tasks do not inherit counters
55-p::
56--pid=<pid>::
57        stat events on existing process id (comma separated list)
58
59-t::
60--tid=<tid>::
61        stat events on existing thread id (comma separated list)
62
63
64-a::
65--all-cpus::
66        system-wide collection from all CPUs (default if no target is specified)
67
68-c::
69--scale::
70	scale/normalize counter values
71
72-d::
73--detailed::
74	print more detailed statistics, can be specified up to 3 times
75
76	   -d:          detailed events, L1 and LLC data cache
77        -d -d:     more detailed events, dTLB and iTLB events
78     -d -d -d:     very detailed events, adding prefetch events
79
80-r::
81--repeat=<n>::
82	repeat command and print average + stddev (max: 100). 0 means forever.
83
84-B::
85--big-num::
86        print large numbers with thousands' separators according to locale
87
88-C::
89--cpu=::
90Count only on the list of CPUs provided. Multiple CPUs can be provided as a
91comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
92In per-thread mode, this option is ignored. The -a option is still necessary
93to activate system-wide monitoring. Default is to count on all CPUs.
94
95-A::
96--no-aggr::
97Do not aggregate counts across all monitored CPUs.
98
99-n::
100--null::
101        null run - don't start any counters
102
103-v::
104--verbose::
105        be more verbose (show counter open errors, etc)
106
107-x SEP::
108--field-separator SEP::
109print counts using a CSV-style output to make it easy to import directly into
110spreadsheets. Columns are separated by the string specified in SEP.
111
112-G name::
113--cgroup name::
114monitor only in the container (cgroup) called "name". This option is available only
115in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
116container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
117can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
118to first event, second cgroup to second event and so on. It is possible to provide
119an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
120corresponding events, i.e., they always refer to events defined earlier on the command
121line.
122
123-o file::
124--output file::
125Print the output into the designated file.
126
127--append::
128Append to the output file designated with the -o option. Ignored if -o is not specified.
129
130--log-fd::
131
132Log output to fd, instead of stderr.  Complementary to --output, and mutually exclusive
133with it.  --append may be used here.  Examples:
134     3>results  perf stat --log-fd 3          -- $cmd
135     3>>results perf stat --log-fd 3 --append -- $cmd
136
137--pre::
138--post::
139	Pre and post measurement hooks, e.g.:
140
141perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
142
143-I msecs::
144--interval-print msecs::
145Print count deltas every N milliseconds (minimum: 10ms)
146The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals.  Use with caution.
147	example: 'perf stat -I 1000 -e cycles -a sleep 5'
148
149--interval-count times::
150Print count deltas for fixed number of times.
151This option should be used together with "-I" option.
152	example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
153
154--timeout msecs::
155Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
156This option is not supported with the "-I" option.
157	example: 'perf stat --time 2000 -e cycles -a'
158
159--metric-only::
160Only print computed metrics. Print them in a single line.
161Don't show any raw values. Not supported with --per-thread.
162
163--per-socket::
164Aggregate counts per processor socket for system-wide mode measurements.  This
165is a useful mode to detect imbalance between sockets.  To enable this mode,
166use --per-socket in addition to -a. (system-wide).  The output includes the
167socket number and the number of online processors on that socket. This is
168useful to gauge the amount of aggregation.
169
170--per-core::
171Aggregate counts per physical processor for system-wide mode measurements.  This
172is a useful mode to detect imbalance between physical cores.  To enable this mode,
173use --per-core in addition to -a. (system-wide).  The output includes the
174core number and the number of online logical processors on that physical processor.
175
176--per-thread::
177Aggregate counts per monitored threads, when monitoring threads (-t option)
178or processes (-p option).
179
180-D msecs::
181--delay msecs::
182After starting the program, wait msecs before measuring. This is useful to
183filter out the startup phase of the program, which is often very different.
184
185-T::
186--transaction::
187
188Print statistics of transactional execution if supported.
189
190STAT RECORD
191-----------
192Stores stat data into perf data file.
193
194-o file::
195--output file::
196Output file name.
197
198STAT REPORT
199-----------
200Reads and reports stat data from perf data file.
201
202-i file::
203--input file::
204Input file name.
205
206--per-socket::
207Aggregate counts per processor socket for system-wide mode measurements.
208
209--per-core::
210Aggregate counts per physical processor for system-wide mode measurements.
211
212-M::
213--metrics::
214Print metrics or metricgroups specified in a comma separated list.
215For a group all metrics from the group are added.
216The events from the metrics are automatically measured.
217See perf list output for the possble metrics and metricgroups.
218
219-A::
220--no-aggr::
221Do not aggregate counts across all monitored CPUs.
222
223--topdown::
224Print top down level 1 metrics if supported by the CPU. This allows to
225determine bottle necks in the CPU pipeline for CPU bound workloads,
226by breaking the cycles consumed down into frontend bound, backend bound,
227bad speculation and retiring.
228
229Frontend bound means that the CPU cannot fetch and decode instructions fast
230enough. Backend bound means that computation or memory access is the bottle
231neck. Bad Speculation means that the CPU wasted cycles due to branch
232mispredictions and similar issues. Retiring means that the CPU computed without
233an apparently bottleneck. The bottleneck is only the real bottleneck
234if the workload is actually bound by the CPU and not by something else.
235
236For best results it is usually a good idea to use it with interval
237mode like -I 1000, as the bottleneck of workloads can change often.
238
239The top down metrics are collected per core instead of per
240CPU thread. Per core mode is automatically enabled
241and -a (global monitoring) is needed, requiring root rights or
242perf.perf_event_paranoid=-1.
243
244Topdown uses the full Performance Monitoring Unit, and needs
245disabling of the NMI watchdog (as root):
246echo 0 > /proc/sys/kernel/nmi_watchdog
247for best results. Otherwise the bottlenecks may be inconsistent
248on workload with changing phases.
249
250This enables --metric-only, unless overriden with --no-metric-only.
251
252To interpret the results it is usually needed to know on which
253CPUs the workload runs on. If needed the CPUs can be forced using
254taskset.
255
256--no-merge::
257Do not merge results from same PMUs.
258
259--smi-cost::
260Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
261
262During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
263freeze core counters on SMI.
264The aperf counter will not be effected by the setting.
265The cost of SMI can be measured by (aperf - unhalted core cycles).
266
267In practice, the percentages of SMI cycles is very useful for performance
268oriented analysis. --metric_only will be applied by default.
269The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
270
271Users who wants to get the actual value can apply --no-metric-only.
272
273EXAMPLES
274--------
275
276$ perf stat -- make -j
277
278 Performance counter stats for 'make -j':
279
280    8117.370256  task clock ticks     #      11.281 CPU utilization factor
281            678  context switches     #       0.000 M/sec
282            133  CPU migrations       #       0.000 M/sec
283         235724  pagefaults           #       0.029 M/sec
284    24821162526  CPU cycles           #    3057.784 M/sec
285    18687303457  instructions         #    2302.138 M/sec
286      172158895  cache references     #      21.209 M/sec
287       27075259  cache misses         #       3.335 M/sec
288
289 Wall-clock time elapsed:   719.554352 msecs
290
291CSV FORMAT
292----------
293
294With -x, perf stat is able to output a not-quite-CSV format output
295Commas in the output are not put into "". To make it easy to parse
296it is recommended to use a different character like -x \;
297
298The fields are in this order:
299
300	- optional usec time stamp in fractions of second (with -I xxx)
301	- optional CPU, core, or socket identifier
302	- optional number of logical CPUs aggregated
303	- counter value
304	- unit of the counter value or empty
305	- event name
306	- run time of counter
307	- percentage of measurement time the counter was running
308	- optional variance if multiple values are collected with -r
309	- optional metric value
310	- optional unit of metric
311
312Additional metrics may be printed with all earlier fields being empty.
313
314SEE ALSO
315--------
316linkperf:perf-top[1], linkperf:perf-list[1]
317