blob: 7e6fb7cbc0f423b09bdb1fb3f5dd243bbe2cd973 [file] [log] [blame]
perf-stat - Run a command and gather performance counter statistics
'perf stat' [-e <EVENT> | --event=EVENT] [-a] <command>
'perf stat' [-e <EVENT> | --event=EVENT] [-a] \-- <command> [<options>]
'perf stat' [-e <EVENT> | --event=EVENT] [-a] record [-o file] \-- <command> [<options>]
'perf stat' report [-i file]
This command runs a command and gathers performance counter statistics
from it.
Any command you can specify in a shell.
Select the PMU event. Selection can be:
- a symbolic event name (use 'perf list' to list all events)
- a raw PMU event (eventsel+umask) in the form of rNNN where NNN is a
hexadecimal event descriptor.
- a symbolic or raw PMU event followed by an optional colon
and a list of event modifiers, e.g., cpu-cycles:p. See the
linkperf:perf-list[1] man page for details on event modifiers.
- a symbolically formed event like 'pmu/param1=0x3,param2/' where
param1 and param2 are defined as formats for the PMU in
'percore' is a event qualifier that sums up the event counts for both
hardware threads in a core. For example:
perf stat -A -a -e cpu/event,percore=1/,otherevent ...
- a symbolically formed event like 'pmu/config=M,config1=N,config2=K/'
where M, N, K are numbers (in decimal, hex, octal format).
Acceptable values for each of 'config', 'config1' and 'config2'
parameters are defined by corresponding entries in
Note that the last two syntaxes support prefix and glob matching in
the PMU name to simplify creation of events across multiple instances
of the same type of PMU in large systems (e.g. memory controller PMUs).
Multiple PMU instances are typical for uncore PMUs, so the prefix
'uncore_' is also ignored when performing this match.
child tasks do not inherit counters
stat events on existing process id (comma separated list)
stat events on existing thread id (comma separated list)
stat events on existing bpf program id (comma separated list),
requiring root rights. bpftool-prog could be used to find program
id all bpf programs in the system. For example:
# bpftool prog | head -n 1
17247: tracepoint name sys_enter tag 192d548b9d754067 gpl
# perf stat -e cycles,instructions --bpf-prog 17247 --timeout 1000
Performance counter stats for 'BPF program(s) 17247':
85,967 cycles
28,982 instructions # 0.34 insn per cycle
1.102235068 seconds time elapsed
Use BPF programs to aggregate readings from perf_events. This
allows multiple perf-stat sessions that are counting the same metric (cycles,
instructions, etc.) to share hardware counters.
To use BPF programs on common events by default, use
"perf config stat.bpf-counter-events=<list_of_events>".
With option "--bpf-counters", different perf-stat sessions share
information about shared BPF programs and maps via a pinned hashmap.
Use "--bpf-attr-map" to specify the path of this pinned hashmap.
The default path is /sys/fs/bpf/perf_attr_map.
--pfm-events events::
Select a PMU event using libpfm4 syntax (see
including support for event filters. For example '--pfm-events
inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
option using the comma separator. Hardware events and generic hardware
events cannot be mixed together. The latter must be used with the -e
option. The -e option and this one can be mixed and matched. Events
can be grouped using the {} notation.
system-wide collection from all CPUs (default if no target is specified)
Don't scale/normalize counter values
print more detailed statistics, can be specified up to 3 times
-d: detailed events, L1 and LLC data cache
-d -d: more detailed events, dTLB and iTLB events
-d -d -d: very detailed events, adding prefetch events
repeat command and print average + stddev (max: 100). 0 means forever.
print large numbers with thousands' separators according to locale.
Enabled by default. Use "--no-big-num" to disable.
Default setting can be changed with "perf config stat.big-num=false".
Count only on the list of CPUs provided. Multiple CPUs can be provided as a
comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
In per-thread mode, this option is ignored. The -a option is still necessary
to activate system-wide monitoring. Default is to count on all CPUs.
Do not aggregate counts across all monitored CPUs.
null run - Don't start any counters.
This can be useful to measure just elapsed wall-clock time - or to assess the
raw overhead of perf stat itself, without running any counters.
be more verbose (show counter open errors, etc)
-x SEP::
--field-separator SEP::
print counts using a CSV-style output to make it easy to import directly into
spreadsheets. Columns are separated by the string specified in SEP.
--table:: Display time for each run (-r option), in a table format, e.g.:
$ perf stat --null -r 5 --table perf bench sched pipe
Performance counter stats for 'perf bench sched pipe' (5 runs):
# Table of individual measurements:
5.189 (-0.293) #
5.189 (-0.294) #
5.186 (-0.296) #
5.663 (+0.181) ##
6.186 (+0.703) ####
# Final result:
5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
-G name::
--cgroup name::
monitor only in the container (cgroup) called "name". This option is available only
in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
to first event, second cgroup to second event and so on. It is possible to provide
an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
corresponding events, i.e., they always refer to events defined earlier on the command
line. If the user wants to track multiple events for a specific cgroup, the user can
use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
--for-each-cgroup name::
Expand event list for each cgroup in "name" (allow multiple cgroups separated
by comma). It also support regex patterns to match multiple groups. This has same
effect that repeating -e option and -G option for each event x name. This option
cannot be used with -G/--cgroup option.
-o file::
--output file::
Print the output into the designated file.
Append to the output file designated with the -o option. Ignored if -o is not specified.
Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
with it. --append may be used here. Examples:
3>results perf stat --log-fd 3 \-- $cmd
3>>results perf stat --log-fd 3 --append \-- $cmd
ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as follows.
Listen on ctl-fd descriptor for command to control measurement ('enable': enable events,
'disable': disable events). Measurements can be started with events disabled using
--delay=-1 option. Optionally send control command completion ('ack\n') to ack-fd descriptor
to synchronize with the controlling process. Example of bash shell script to enable and
disable events during measurements:
test -p ${ctl_fifo} && unlink ${ctl_fifo}
mkfifo ${ctl_fifo}
exec {ctl_fd}<>${ctl_fifo}
test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
mkfifo ${ctl_ack_fifo}
exec {ctl_fd_ack}<>${ctl_ack_fifo}
perf stat -D -1 -e cpu-cycles -a -I 1000 \
--control fd:${ctl_fd},${ctl_fd_ack} \
\-- sleep 30 &
sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
exec {ctl_fd_ack}>&-
unlink ${ctl_ack_fifo}
exec {ctl_fd}>&-
unlink ${ctl_fifo}
wait -n ${perf_pid}
exit $?
Pre and post measurement hooks, e.g.:
perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' \-- make -s -j64 O=defconfig-build/ bzImage
-I msecs::
--interval-print msecs::
Print count deltas every N milliseconds (minimum: 1ms)
The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
example: 'perf stat -I 1000 -e cycles -a sleep 5'
If the metric exists, it is calculated by the counts generated in this interval and the metric is printed after #.
--interval-count times::
Print count deltas for fixed number of times.
This option should be used together with "-I" option.
example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
Clear the screen before next interval.
--timeout msecs::
Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
This option is not supported with the "-I" option.
example: 'perf stat --time 2000 -e cycles -a'
Only print computed metrics. Print them in a single line.
Don't show any raw values. Not supported with --per-thread.
Aggregate counts per processor socket for system-wide mode measurements. This
is a useful mode to detect imbalance between sockets. To enable this mode,
use --per-socket in addition to -a. (system-wide). The output includes the
socket number and the number of online processors on that socket. This is
useful to gauge the amount of aggregation.
Aggregate counts per processor die for system-wide mode measurements. This
is a useful mode to detect imbalance between dies. To enable this mode,
use --per-die in addition to -a. (system-wide). The output includes the
die number and the number of online processors on that die. This is
useful to gauge the amount of aggregation.
Aggregate counts per physical processor for system-wide mode measurements. This
is a useful mode to detect imbalance between physical cores. To enable this mode,
use --per-core in addition to -a. (system-wide). The output includes the
core number and the number of online logical processors on that physical processor.
Aggregate counts per monitored threads, when monitoring threads (-t option)
or processes (-p option).
Aggregate counts per NUMA nodes for system-wide mode measurements. This
is a useful mode to detect imbalance between NUMA nodes. To enable this
mode, use --per-node in addition to -a. (system-wide).
-D msecs::
--delay msecs::
After starting the program, wait msecs before measuring (-1: start with events
disabled). This is useful to filter out the startup phase of the program,
which is often very different.
Print statistics of transactional execution if supported.
By default, events to compute a metric are placed in weak groups. The
group tries to enforce scheduling all or none of the events. The
--metric-no-group option places events outside of groups and may
increase the chance of the event being scheduled - leading to more
accuracy. However, as events may not be scheduled together accuracy
for metrics like instructions per cycle can be lower - as both metrics
may no longer be being measured at the same time.
By default metric events in different weak groups can be shared if one
group contains all the events needed by another. In such cases one
group will be eliminated reducing event multiplexing and making it so
that certain groups of metrics sum to 100%. A downside to sharing a
group is that the group may require multiplexing and so accuracy for a
small group that need not have multiplexing is lowered. This option
forbids the event merging logic from sharing events between groups and
may be used to increase accuracy in this case.
Don't print output. This is useful with perf stat record below to only
write data to the file.
Stores stat data into perf data file.
-o file::
--output file::
Output file name.
Reads and reports stat data from perf data file.
-i file::
--input file::
Input file name.
Aggregate counts per processor socket for system-wide mode measurements.
Aggregate counts per processor die for system-wide mode measurements.
Aggregate counts per physical processor for system-wide mode measurements.
Print metrics or metricgroups specified in a comma separated list.
For a group all metrics from the group are added.
The events from the metrics are automatically measured.
See perf list output for the possible metrics and metricgroups.
Do not aggregate counts across all monitored CPUs.
Print complete top-down metrics supported by the CPU. This allows to
determine bottle necks in the CPU pipeline for CPU bound workloads,
by breaking the cycles consumed down into frontend bound, backend bound,
bad speculation and retiring.
Frontend bound means that the CPU cannot fetch and decode instructions fast
enough. Backend bound means that computation or memory access is the bottle
neck. Bad Speculation means that the CPU wasted cycles due to branch
mispredictions and similar issues. Retiring means that the CPU computed without
an apparently bottleneck. The bottleneck is only the real bottleneck
if the workload is actually bound by the CPU and not by something else.
For best results it is usually a good idea to use it with interval
mode like -I 1000, as the bottleneck of workloads can change often.
This enables --metric-only, unless overridden with --no-metric-only.
The following restrictions only apply to older Intel CPUs and Atom,
on newer CPUs (IceLake and later) TopDown can be collected for any thread:
The top down metrics are collected per core instead of per
CPU thread. Per core mode is automatically enabled
and -a (global monitoring) is needed, requiring root rights or
Topdown uses the full Performance Monitoring Unit, and needs
disabling of the NMI watchdog (as root):
echo 0 > /proc/sys/kernel/nmi_watchdog
for best results. Otherwise the bottlenecks may be inconsistent
on workload with changing phases.
To interpret the results it is usually needed to know on which
CPUs the workload runs on. If needed the CPUs can be forced using
Print the top-down statistics that equal to or lower than the input level.
It allows users to print the interested top-down metrics level instead of
the complete top-down metrics.
The availability of the top-down metrics level depends on the hardware. For
example, Ice Lake only supports L1 top-down metrics. The Sapphire Rapids
supports both L1 and L2 top-down metrics.
Default: 0 means the max level that the current hardware support.
Error out if the input is higher than the supported max level.
Do not merge results from same PMUs.
When multiple events are created from a single event specification,
stat will, by default, aggregate the event counts and show the result
in a single row. This option disables that behavior and shows
the individual events and counts.
Multiple events are created from a single event specification when:
1. Prefix or glob matching is used for the PMU name.
2. Aliases, which are listed immediately after the Kernel PMU events
by perf list, are used.
Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
freeze core counters on SMI.
The aperf counter will not be effected by the setting.
The cost of SMI can be measured by (aperf - unhalted core cycles).
In practice, the percentages of SMI cycles is very useful for performance
oriented analysis. --metric_only will be applied by default.
The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
Users who wants to get the actual value can apply --no-metric-only.
Configure all used events to run in kernel space.
Configure all used events to run in user space.
The event modifier "percore" has supported to sum up the event counts
for all hardware threads in a core and show the counts per core.
This option with event modifier "percore" enabled also sums up the event
counts for all hardware threads in a core but show the sum counts per
hardware thread. This is essentially a replacement for the any bit and
convenient for post processing.
Print summary for interval mode (-I).
Don't print 'summary' at the first column for CVS summary output.
This option must be used with -x and --summary.
This option can be enabled in perf config by setting the variable
$ perf config
$ perf stat \-- make
Performance counter stats for 'make':
83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
3,228,188 page-faults:u # 0.039 M/sec
229,570,665,834 cycles:u # 2.742 GHz
313,163,853,778 instructions:u # 1.36 insn per cycle
69,704,684,856 branches:u # 832.559 M/sec
2,078,861,393 branch-misses:u # 2.98% of all branches
83.409183620 seconds time elapsed
74.684747000 seconds user
8.739217000 seconds sys
As displayed in the example above we can display 3 types of timings.
We always display the time the counters were enabled/alive:
83.409183620 seconds time elapsed
For workload sessions we also display time the workloads spent in
user/system lands:
74.684747000 seconds user
8.739217000 seconds sys
Those times are the very same as displayed by the 'time' tool.
With -x, perf stat is able to output a not-quite-CSV format output
Commas in the output are not put into "". To make it easy to parse
it is recommended to use a different character like -x \;
The fields are in this order:
- optional usec time stamp in fractions of second (with -I xxx)
- optional CPU, core, or socket identifier
- optional number of logical CPUs aggregated
- counter value
- unit of the counter value or empty
- event name
- run time of counter
- percentage of measurement time the counter was running
- optional variance if multiple values are collected with -r
- optional metric value
- optional unit of metric
Additional metrics may be printed with all earlier fields being empty.
linkperf:perf-top[1], linkperf:perf-list[1]