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 'percore' is a event qualifier that sums up the event counts for both 47 hardware threads in a core. For example: 48 perf stat -A -a -e cpu/event,percore=1/,otherevent ... 49 50 - a symbolically formed event like 'pmu/config=M,config1=N,config2=K/' 51 where M, N, K are numbers (in decimal, hex, octal format). 52 Acceptable values for each of 'config', 'config1' and 'config2' 53 parameters are defined by corresponding entries in 54 /sys/bus/event_source/devices/<pmu>/format/* 55 56 Note that the last two syntaxes support prefix and glob matching in 57 the PMU name to simplify creation of events across multiple instances 58 of the same type of PMU in large systems (e.g. memory controller PMUs). 59 Multiple PMU instances are typical for uncore PMUs, so the prefix 60 'uncore_' is also ignored when performing this match. 61 62 63-i:: 64--no-inherit:: 65 child tasks do not inherit counters 66-p:: 67--pid=<pid>:: 68 stat events on existing process id (comma separated list) 69 70-t:: 71--tid=<tid>:: 72 stat events on existing thread id (comma separated list) 73 74ifdef::HAVE_LIBPFM[] 75--pfm-events events:: 76Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net) 77including support for event filters. For example '--pfm-events 78inst_retired:any_p:u:c=1:i'. More than one event can be passed to the 79option using the comma separator. Hardware events and generic hardware 80events cannot be mixed together. The latter must be used with the -e 81option. The -e option and this one can be mixed and matched. Events 82can be grouped using the {} notation. 83endif::HAVE_LIBPFM[] 84 85-a:: 86--all-cpus:: 87 system-wide collection from all CPUs (default if no target is specified) 88 89--no-scale:: 90 Don't scale/normalize counter values 91 92-d:: 93--detailed:: 94 print more detailed statistics, can be specified up to 3 times 95 96 -d: detailed events, L1 and LLC data cache 97 -d -d: more detailed events, dTLB and iTLB events 98 -d -d -d: very detailed events, adding prefetch events 99 100-r:: 101--repeat=<n>:: 102 repeat command and print average + stddev (max: 100). 0 means forever. 103 104-B:: 105--big-num:: 106 print large numbers with thousands' separators according to locale. 107 Enabled by default. Use "--no-big-num" to disable. 108 Default setting can be changed with "perf config stat.big-num=false". 109 110-C:: 111--cpu=:: 112Count only on the list of CPUs provided. Multiple CPUs can be provided as a 113comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2. 114In per-thread mode, this option is ignored. The -a option is still necessary 115to activate system-wide monitoring. Default is to count on all CPUs. 116 117-A:: 118--no-aggr:: 119Do not aggregate counts across all monitored CPUs. 120 121-n:: 122--null:: 123 null run - don't start any counters 124 125-v:: 126--verbose:: 127 be more verbose (show counter open errors, etc) 128 129-x SEP:: 130--field-separator SEP:: 131print counts using a CSV-style output to make it easy to import directly into 132spreadsheets. Columns are separated by the string specified in SEP. 133 134--table:: Display time for each run (-r option), in a table format, e.g.: 135 136 $ perf stat --null -r 5 --table perf bench sched pipe 137 138 Performance counter stats for 'perf bench sched pipe' (5 runs): 139 140 # Table of individual measurements: 141 5.189 (-0.293) # 142 5.189 (-0.294) # 143 5.186 (-0.296) # 144 5.663 (+0.181) ## 145 6.186 (+0.703) #### 146 147 # Final result: 148 5.483 +- 0.198 seconds time elapsed ( +- 3.62% ) 149 150-G name:: 151--cgroup name:: 152monitor only in the container (cgroup) called "name". This option is available only 153in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to 154container "name" are monitored when they run on the monitored CPUs. Multiple cgroups 155can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup 156to first event, second cgroup to second event and so on. It is possible to provide 157an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have 158corresponding events, i.e., they always refer to events defined earlier on the command 159line. If the user wants to track multiple events for a specific cgroup, the user can 160use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'. 161 162If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this 163command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'. 164 165-o file:: 166--output file:: 167Print the output into the designated file. 168 169--append:: 170Append to the output file designated with the -o option. Ignored if -o is not specified. 171 172--log-fd:: 173 174Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive 175with it. --append may be used here. Examples: 176 3>results perf stat --log-fd 3 -- $cmd 177 3>>results perf stat --log-fd 3 --append -- $cmd 178 179--pre:: 180--post:: 181 Pre and post measurement hooks, e.g.: 182 183perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage 184 185-I msecs:: 186--interval-print msecs:: 187Print count deltas every N milliseconds (minimum: 1ms) 188The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution. 189 example: 'perf stat -I 1000 -e cycles -a sleep 5' 190 191If the metric exists, it is calculated by the counts generated in this interval and the metric is printed after #. 192 193--interval-count times:: 194Print count deltas for fixed number of times. 195This option should be used together with "-I" option. 196 example: 'perf stat -I 1000 --interval-count 2 -e cycles -a' 197 198--interval-clear:: 199Clear the screen before next interval. 200 201--timeout msecs:: 202Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms). 203This option is not supported with the "-I" option. 204 example: 'perf stat --time 2000 -e cycles -a' 205 206--metric-only:: 207Only print computed metrics. Print them in a single line. 208Don't show any raw values. Not supported with --per-thread. 209 210--per-socket:: 211Aggregate counts per processor socket for system-wide mode measurements. This 212is a useful mode to detect imbalance between sockets. To enable this mode, 213use --per-socket in addition to -a. (system-wide). The output includes the 214socket number and the number of online processors on that socket. This is 215useful to gauge the amount of aggregation. 216 217--per-die:: 218Aggregate counts per processor die for system-wide mode measurements. This 219is a useful mode to detect imbalance between dies. To enable this mode, 220use --per-die in addition to -a. (system-wide). The output includes the 221die number and the number of online processors on that die. This is 222useful to gauge the amount of aggregation. 223 224--per-core:: 225Aggregate counts per physical processor for system-wide mode measurements. This 226is a useful mode to detect imbalance between physical cores. To enable this mode, 227use --per-core in addition to -a. (system-wide). The output includes the 228core number and the number of online logical processors on that physical processor. 229 230--per-thread:: 231Aggregate counts per monitored threads, when monitoring threads (-t option) 232or processes (-p option). 233 234--per-node:: 235Aggregate counts per NUMA nodes for system-wide mode measurements. This 236is a useful mode to detect imbalance between NUMA nodes. To enable this 237mode, use --per-node in addition to -a. (system-wide). 238 239-D msecs:: 240--delay msecs:: 241After starting the program, wait msecs before measuring. This is useful to 242filter out the startup phase of the program, which is often very different. 243 244-T:: 245--transaction:: 246 247Print statistics of transactional execution if supported. 248 249--metric-no-group:: 250By default, events to compute a metric are placed in weak groups. The 251group tries to enforce scheduling all or none of the events. The 252--metric-no-group option places events outside of groups and may 253increase the chance of the event being scheduled - leading to more 254accuracy. However, as events may not be scheduled together accuracy 255for metrics like instructions per cycle can be lower - as both metrics 256may no longer be being measured at the same time. 257 258--metric-no-merge:: 259By default metric events in different weak groups can be shared if one 260group contains all the events needed by another. In such cases one 261group will be eliminated reducing event multiplexing and making it so 262that certain groups of metrics sum to 100%. A downside to sharing a 263group is that the group may require multiplexing and so accuracy for a 264small group that need not have multiplexing is lowered. This option 265forbids the event merging logic from sharing events between groups and 266may be used to increase accuracy in this case. 267 268STAT RECORD 269----------- 270Stores stat data into perf data file. 271 272-o file:: 273--output file:: 274Output file name. 275 276STAT REPORT 277----------- 278Reads and reports stat data from perf data file. 279 280-i file:: 281--input file:: 282Input file name. 283 284--per-socket:: 285Aggregate counts per processor socket for system-wide mode measurements. 286 287--per-die:: 288Aggregate counts per processor die for system-wide mode measurements. 289 290--per-core:: 291Aggregate counts per physical processor for system-wide mode measurements. 292 293-M:: 294--metrics:: 295Print metrics or metricgroups specified in a comma separated list. 296For a group all metrics from the group are added. 297The events from the metrics are automatically measured. 298See perf list output for the possble metrics and metricgroups. 299 300-A:: 301--no-aggr:: 302Do not aggregate counts across all monitored CPUs. 303 304--topdown:: 305Print top down level 1 metrics if supported by the CPU. This allows to 306determine bottle necks in the CPU pipeline for CPU bound workloads, 307by breaking the cycles consumed down into frontend bound, backend bound, 308bad speculation and retiring. 309 310Frontend bound means that the CPU cannot fetch and decode instructions fast 311enough. Backend bound means that computation or memory access is the bottle 312neck. Bad Speculation means that the CPU wasted cycles due to branch 313mispredictions and similar issues. Retiring means that the CPU computed without 314an apparently bottleneck. The bottleneck is only the real bottleneck 315if the workload is actually bound by the CPU and not by something else. 316 317For best results it is usually a good idea to use it with interval 318mode like -I 1000, as the bottleneck of workloads can change often. 319 320The top down metrics are collected per core instead of per 321CPU thread. Per core mode is automatically enabled 322and -a (global monitoring) is needed, requiring root rights or 323perf.perf_event_paranoid=-1. 324 325Topdown uses the full Performance Monitoring Unit, and needs 326disabling of the NMI watchdog (as root): 327echo 0 > /proc/sys/kernel/nmi_watchdog 328for best results. Otherwise the bottlenecks may be inconsistent 329on workload with changing phases. 330 331This enables --metric-only, unless overridden with --no-metric-only. 332 333To interpret the results it is usually needed to know on which 334CPUs the workload runs on. If needed the CPUs can be forced using 335taskset. 336 337--no-merge:: 338Do not merge results from same PMUs. 339 340When multiple events are created from a single event specification, 341stat will, by default, aggregate the event counts and show the result 342in a single row. This option disables that behavior and shows 343the individual events and counts. 344 345Multiple events are created from a single event specification when: 3461. Prefix or glob matching is used for the PMU name. 3472. Aliases, which are listed immediately after the Kernel PMU events 348 by perf list, are used. 349 350--smi-cost:: 351Measure SMI cost if msr/aperf/ and msr/smi/ events are supported. 352 353During the measurement, the /sys/device/cpu/freeze_on_smi will be set to 354freeze core counters on SMI. 355The aperf counter will not be effected by the setting. 356The cost of SMI can be measured by (aperf - unhalted core cycles). 357 358In practice, the percentages of SMI cycles is very useful for performance 359oriented analysis. --metric_only will be applied by default. 360The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf 361 362Users who wants to get the actual value can apply --no-metric-only. 363 364--all-kernel:: 365Configure all used events to run in kernel space. 366 367--all-user:: 368Configure all used events to run in user space. 369 370--percore-show-thread:: 371The event modifier "percore" has supported to sum up the event counts 372for all hardware threads in a core and show the counts per core. 373 374This option with event modifier "percore" enabled also sums up the event 375counts for all hardware threads in a core but show the sum counts per 376hardware thread. This is essentially a replacement for the any bit and 377convenient for post processing. 378 379EXAMPLES 380-------- 381 382$ perf stat -- make 383 384 Performance counter stats for 'make': 385 386 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized 387 0 context-switches:u # 0.000 K/sec 388 0 cpu-migrations:u # 0.000 K/sec 389 3,228,188 page-faults:u # 0.039 M/sec 390 229,570,665,834 cycles:u # 2.742 GHz 391 313,163,853,778 instructions:u # 1.36 insn per cycle 392 69,704,684,856 branches:u # 832.559 M/sec 393 2,078,861,393 branch-misses:u # 2.98% of all branches 394 395 83.409183620 seconds time elapsed 396 397 74.684747000 seconds user 398 8.739217000 seconds sys 399 400TIMINGS 401------- 402As displayed in the example above we can display 3 types of timings. 403We always display the time the counters were enabled/alive: 404 405 83.409183620 seconds time elapsed 406 407For workload sessions we also display time the workloads spent in 408user/system lands: 409 410 74.684747000 seconds user 411 8.739217000 seconds sys 412 413Those times are the very same as displayed by the 'time' tool. 414 415CSV FORMAT 416---------- 417 418With -x, perf stat is able to output a not-quite-CSV format output 419Commas in the output are not put into "". To make it easy to parse 420it is recommended to use a different character like -x \; 421 422The fields are in this order: 423 424 - optional usec time stamp in fractions of second (with -I xxx) 425 - optional CPU, core, or socket identifier 426 - optional number of logical CPUs aggregated 427 - counter value 428 - unit of the counter value or empty 429 - event name 430 - run time of counter 431 - percentage of measurement time the counter was running 432 - optional variance if multiple values are collected with -r 433 - optional metric value 434 - optional unit of metric 435 436Additional metrics may be printed with all earlier fields being empty. 437 438SEE ALSO 439-------- 440linkperf:perf-top[1], linkperf:perf-list[1] 441