kernel.git/tools/perf/Documentation/perf-bench.txt, branch master

perf bench mem: Add mmap() workloads

2025-09-19T15:43:59Z

Add two mmap() workloads: one that eagerly populates a region and another that demand faults it in. The intent is to probe the memory subsytem performance incurred by mmap(). $ perf bench mem mmap -s 4gb -p 4kb -l 10 -f populate # Running 'mem/mmap' benchmark: # function 'populate' (Eagerly populated map()) # Copying 4gb bytes ... 1.811691 GB/sec $ perf bench mem mmap -s 4gb -p 2mb -l 10 -f populate # Running 'mem/mmap' benchmark: # function 'populate' (Eagerly populated mmap()) # Copying 4gb bytes ... 12.272017 GB/sec $ perf bench mem mmap -s 4gb -p 1gb -l 10 -f populate # Running 'mem/mmap' benchmark: # function 'populate' (Eagerly populated mmap()) # Copying 4gb bytes ... 17.085927 GB/sec Signed-off-by: Ankur Arora Tested-by: Arnaldo Carvalho de Melo Cc: Andrew Morton Cc: Andy Lutomirski Cc: Boris Ostrovsky Cc: Borislav Petkov Cc: Dave Hansen Cc: David Hildenbrand Cc: H. Peter Anvin Cc: Ingo Molnar Cc: Konrad Rzeszutek Wilk Cc: Mateusz Guzik Cc: Matthew Wilcox Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Raghavendra K T Cc: Thomas Gleixner Signed-off-by: Arnaldo Carvalho de Melo

perf bench mem: Allow chunking on a memory region

2025-09-19T15:43:38Z

There can be a significant gap in memset/memcpy performance depending on the size of the region being operated on. With chunk-size=4kb: $ echo madvise > /sys/kernel/mm/transparent_hugepage/enabled $ perf bench mem memset -p 4kb -k 4kb -s 4gb -l 10 -f x86-64-stosq # Running 'mem/memset' benchmark: # function 'x86-64-stosq' (movsq-based memset() in arch/x86/lib/memset_64.S) # Copying 4gb bytes ... 13.011655 GB/sec With chunk-size=1gb: $ echo madvise > /sys/kernel/mm/transparent_hugepage/enabled $ perf bench mem memset -p 4kb -k 1gb -s 4gb -l 10 -f x86-64-stosq # Running 'mem/memset' benchmark: # function 'x86-64-stosq' (movsq-based memset() in arch/x86/lib/memset_64.S) # Copying 4gb bytes ... 21.936355 GB/sec So, allow the user to specify the chunk-size. The default value is identical to the total size of the region, which preserves current behaviour. Reviewed-by: Namhyung Kim Signed-off-by: Ankur Arora Tested-by: Arnaldo Carvalho de Melo Cc: Andrew Morton Cc: Andy Lutomirski Cc: Boris Ostrovsky Cc: Borislav Petkov Cc: Dave Hansen Cc: David Hildenbrand Cc: H. Peter Anvin Cc: Ingo Molnar Cc: Konrad Rzeszutek Wilk Cc: Mateusz Guzik Cc: Matthew Wilcox Cc: Peter Zijlstra Cc: Raghavendra K T Cc: Thomas Gleixner Signed-off-by: Arnaldo Carvalho de Melo

perf bench mem: Allow mapping of hugepages

2025-09-19T15:43:26Z

Page sizes that can be selected: 4KB, 2MB, 1GB. Both the reservation and node from which hugepages are allocated from are expected to be addressed by the user. An example of page-size selection: $ perf bench mem memset -s 4gb -p 2mb # Running 'mem/memset' benchmark: # function 'default' (Default memset() provided by glibc) # Copying 4gb bytes ... 14.919194 GB/sec # function 'x86-64-unrolled' (unrolled memset() in arch/x86/lib/memset_64.S) # Copying 4gb bytes ... 11.514503 GB/sec # function 'x86-64-stosq' (movsq-based memset() in arch/x86/lib/memset_64.S) # Copying 4gb bytes ... 12.600568 GB/sec Signed-off-by: Ankur Arora Tested-by: Arnaldo Carvalho de Melo Cc: Andrew Morton Cc: Andy Lutomirski Cc: Boris Ostrovsky Cc: Borislav Petkov Cc: Dave Hansen Cc: David Hildenbrand Cc: H. Peter Anvin Cc: Ingo Molnar Cc: Konrad Rzeszutek Wilk Cc: Mateusz Guzik Cc: Matthew Wilcox Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Raghavendra K T Cc: Thomas Gleixner Signed-off-by: Arnaldo Carvalho de Melo

perf bench sched pipe: Add -G/--cgroups option

2023-10-25T17:02:10Z

The -G/--cgroups option is to put sender and receiver in different cgroups in order to measure cgroup context switch overheads. Users need to make sure the cgroups exist and accessible. The following example should the effect of this change. Please don't forget taskset before the perf bench to measure cgroup switches properly. Otherwise each task would run on a different CPU and generate cgroup switches regardless of this change. # perf stat -e context-switches,cgroup-switches \ > taskset -c 0 perf bench sched pipe -l 10000 > /dev/null Performance counter stats for 'taskset -c 0 perf bench sched pipe -l 10000': 20,001 context-switches 2 cgroup-switches 0.053449651 seconds time elapsed 0.011286000 seconds user 0.041869000 seconds sys # perf stat -e context-switches,cgroup-switches \ > taskset -c 0 perf bench sched pipe -l 10000 -G AAA,BBB > /dev/null Performance counter stats for 'taskset -c 0 perf bench sched pipe -l 10000 -G AAA,BBB': 20,001 context-switches 20,001 cgroup-switches 0.052768627 seconds time elapsed 0.006284000 seconds user 0.046266000 seconds sys Signed-off-by: Namhyung Kim Tested-by: Arnaldo Carvalho de Melo Link: https://lore.kernel.org/r/20231017202342.1353124-1-namhyung@kernel.org

perf bench uprobe: Add benchmark to test uprobe overhead

2023-07-20T14:31:19Z

This just adds the initial "workload", a call to libc's usleep(1000us) function: $ perf stat --null perf bench uprobe all # Running uprobe/baseline benchmark... # Executed 1000 usleep(1000) calls Total time: 1053533 usecs 1053.533 usecs/op Performance counter stats for 'perf bench uprobe all': 1.061042896 seconds time elapsed 0.001079000 seconds user 0.006499000 seconds sys $ More entries will be added using a BPF skel to add various uprobes to the usleep() function. Acked-by: Ian Rogers Cc: Adrian Hunter Cc: Andre Fredette Cc: Clark Williams Cc: Dave Tucker Cc: Derek Barbosa Cc: Jiri Olsa Cc: Masami Hiramatsu (Google) Cc: Namhyung Kim Link: https://lore.kernel.org/lkml/20230719204910.539044-2-acme@kernel.org Signed-off-by: Arnaldo Carvalho de Melo

perf: Various spelling fixes

2023-01-23T13:00:47Z

Fix various spelling errors as reported by Debian's lintian tool. "amount of times" -> "number of times" ocurrence -> occurrence upto -> up to Signed-off-by: Diederik de Haas Acked-by: Ian Rogers Cc: Alexander Shishkin Cc: Ingo Molnar Cc: Jiri Olsa Cc: Mark Rutland Cc: Namhyung Kim Cc: Peter Zijlstra Link: https://lore.kernel.org/r/20230122122034.48020-1-didi.debian@cknow.org Signed-off-by: Arnaldo Carvalho de Melo

perf bench: Add basic syscall benchmark

2020-07-28T11:50:48Z

The usefulness of having a standard way of testing syscall performance has come up from time to time[0]. Furthermore, some of our testing machinery (such as 'mmtests') already makes use of a simplified version of the microbenchmark. This patch mainly takes the same idea to measure syscall throughput compatible with 'perf-bench' via getppid(2), yet without any of the additional template stuff from Ingo's version (based on numa.c). The code is identical to what mmtests uses. [0] https://lore.kernel.org/lkml/20160201074156.GA27156@gmail.com/ Committer notes: Add mising stdlib.h and unistd.h to get the prototypes for exit() and getppid(). Committer testing: $ perf bench Usage: perf bench [] [] # List of all available benchmark collections: sched: Scheduler and IPC benchmarks syscall: System call benchmarks mem: Memory access benchmarks numa: NUMA scheduling and MM benchmarks futex: Futex stressing benchmarks epoll: Epoll stressing benchmarks internals: Perf-internals benchmarks all: All benchmarks $ $ perf bench syscall # List of available benchmarks for collection 'syscall': basic: Benchmark for basic getppid(2) calls all: Run all syscall benchmarks $ perf bench syscall basic # Running 'syscall/basic' benchmark: # Executed 10000000 getppid() calls Total time: 3.679 [sec] 0.367957 usecs/op 2717708 ops/sec $ perf bench syscall all # Running syscall/basic benchmark... # Executed 10000000 getppid() calls Total time: 3.644 [sec] 0.364456 usecs/op 2743815 ops/sec $ Signed-off-by: Davidlohr Bueso Acked-by: Josh Poimboeuf Acked-by: Mel Gorman Tested-by: Arnaldo Carvalho de Melo Cc: Jiri Olsa Cc: Namhyung Kim Link: http://lore.kernel.org/lkml/20190308181747.l36zqz2avtivrr3c@linux-r8p5 Signed-off-by: Arnaldo Carvalho de Melo

perf bench: Add event synthesis benchmark

2020-04-16T15:19:12Z

Event synthesis may occur at the start or end (tail) of a perf command. In system-wide mode it can scan every process in /proc, which may add seconds of latency before event recording. Add a new benchmark that times how long event synthesis takes with and without data synthesis. An example execution looks like: $ perf bench internals synthesize # Running 'internals/synthesize' benchmark: Average synthesis took: 168.253800 usec Average data synthesis took: 208.104700 usec Signed-off-by: Ian Rogers Acked-by: Jiri Olsa Tested-by: Arnaldo Carvalho de Melo Cc: Alexander Shishkin Cc: Andrey Zhizhikin Cc: Kan Liang Cc: Kefeng Wang Cc: Mark Rutland Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Petr Mladek Cc: Stephane Eranian Cc: Thomas Gleixner Link: http://lore.kernel.org/lkml/20200402154357.107873-2-irogers@google.com Signed-off-by: Arnaldo Carvalho de Melo

perf bench: Add epoll_ctl(2) benchmark

2018-11-22T01:39:55Z

Benchmark the various operations allowed for epoll_ctl(2). The idea is to concurrently stress a single epoll instance doing add/mod/del operations. Committer testing: # perf bench epoll ctl # Running 'epoll/ctl' benchmark: Run summary [PID 20344]: 4 threads doing epoll_ctl ops 64 file-descriptors for 8 secs. [thread 0] fdmap: 0x21a46b0 ... 0x21a47ac [ add: 1680960 ops; mod: 1680960 ops; del: 1680960 ops ] [thread 1] fdmap: 0x21a4960 ... 0x21a4a5c [ add: 1685440 ops; mod: 1685440 ops; del: 1685440 ops ] [thread 2] fdmap: 0x21a4c10 ... 0x21a4d0c [ add: 1674368 ops; mod: 1674368 ops; del: 1674368 ops ] [thread 3] fdmap: 0x21a4ec0 ... 0x21a4fbc [ add: 1677568 ops; mod: 1677568 ops; del: 1677568 ops ] Averaged 1679584 ADD operations (+- 0.14%) Averaged 1679584 MOD operations (+- 0.14%) Averaged 1679584 DEL operations (+- 0.14%) # Lets measure those calls with 'perf trace' to get a glympse at what this benchmark is doing in terms of syscalls: # perf trace -m32768 -s perf bench epoll ctl # Running 'epoll/ctl' benchmark: Run summary [PID 20405]: 4 threads doing epoll_ctl ops 64 file-descriptors for 8 secs. [thread 0] fdmap: 0x21764e0 ... 0x21765dc [ add: 1100480 ops; mod: 1100480 ops; del: 1100480 ops ] [thread 1] fdmap: 0x2176790 ... 0x217688c [ add: 1250176 ops; mod: 1250176 ops; del: 1250176 ops ] [thread 2] fdmap: 0x2176a40 ... 0x2176b3c [ add: 1022464 ops; mod: 1022464 ops; del: 1022464 ops ] [thread 3] fdmap: 0x2176cf0 ... 0x2176dec [ add: 705472 ops; mod: 705472 ops; del: 705472 ops ] Averaged 1019648 ADD operations (+- 11.27%) Averaged 1019648 MOD operations (+- 11.27%) Averaged 1019648 DEL operations (+- 11.27%) Summary of events: epoll-ctl (20405), 1264 events, 0.0% syscall calls total min avg max stddev (msec) (msec) (msec) (msec) (%) --------------- -------- --------- --------- --------- --------- ------ eventfd2 256 9.514 0.001 0.037 5.243 68.00% clone 4 1.245 0.204 0.311 0.531 24.13% mprotect 66 0.345 0.002 0.005 0.021 7.43% openat 45 0.313 0.004 0.007 0.073 21.93% mmap 88 0.302 0.002 0.003 0.013 5.02% futex 4 0.160 0.002 0.040 0.140 83.43% sched_setaffinity 4 0.124 0.005 0.031 0.070 49.39% read 44 0.103 0.001 0.002 0.013 15.54% fstat 40 0.052 0.001 0.001 0.003 5.43% close 39 0.039 0.001 0.001 0.001 1.48% stat 9 0.034 0.003 0.004 0.006 7.30% access 3 0.023 0.007 0.008 0.008 4.25% open 2 0.021 0.008 0.011 0.013 22.60% getdents 4 0.019 0.001 0.005 0.009 37.15% write 2 0.013 0.004 0.007 0.009 38.48% munmap 1 0.010 0.010 0.010 0.010 0.00% brk 3 0.006 0.001 0.002 0.003 26.34% rt_sigprocmask 2 0.004 0.001 0.002 0.003 43.95% rt_sigaction 3 0.004 0.001 0.001 0.002 16.07% prlimit64 3 0.004 0.001 0.001 0.001 5.39% prctl 1 0.003 0.003 0.003 0.003 0.00% epoll_create 1 0.003 0.003 0.003 0.003 0.00% lseek 2 0.002 0.001 0.001 0.001 11.42% sched_getaffinity 1 0.002 0.002 0.002 0.002 0.00% arch_prctl 1 0.002 0.002 0.002 0.002 0.00% set_tid_address 1 0.001 0.001 0.001 0.001 0.00% getpid 1 0.001 0.001 0.001 0.001 0.00% set_robust_list 1 0.001 0.001 0.001 0.001 0.00% execve 1 0.000 0.000 0.000 0.000 0.00% epoll-ctl (20406), 1245480 events, 14.6% syscall calls total min avg max stddev (msec) (msec) (msec) (msec) (%) --------------- -------- --------- --------- --------- --------- ------ epoll_ctl 619511 1034.927 0.001 0.002 6.691 0.67% nanosleep 3226 616.114 0.006 0.191 10.376 7.57% futex 2 11.336 0.002 5.668 11.334 99.97% set_robust_list 1 0.001 0.001 0.001 0.001 0.00% clone 1 0.000 0.000 0.000 0.000 0.00% epoll-ctl (20407), 1243151 events, 14.5% syscall calls total min avg max stddev (msec) (msec) (msec) (msec) (%) --------------- -------- --------- --------- --------- --------- ------ epoll_ctl 618350 1042.181 0.001 0.002 2.512 0.40% nanosleep 3220 366.261 0.012 0.114 18.162 9.59% futex 4 5.463 0.001 1.366 5.427 99.12% set_robust_list 1 0.002 0.002 0.002 0.002 0.00% epoll-ctl (20408), 1801690 events, 21.1% syscall calls total min avg max stddev (msec) (msec) (msec) (msec) (%) --------------- -------- --------- --------- --------- --------- ------ epoll_ctl 896174 1540.581 0.001 0.002 6.987 0.74% nanosleep 4667 783.393 0.006 0.168 10.419 7.10% futex 2 4.682 0.002 2.341 4.681 99.93% set_robust_list 1 0.002 0.002 0.002 0.002 0.00% clone 1 0.000 0.000 0.000 0.000 0.00% epoll-ctl (20409), 4254890 events, 49.8% syscall calls total min avg max stddev (msec) (msec) (msec) (msec) (%) --------------- -------- --------- --------- --------- --------- ------ epoll_ctl 2116416 3768.097 0.001 0.002 9.956 0.41% nanosleep 11023 1141.778 0.006 0.104 9.447 4.95% futex 3 0.037 0.002 0.012 0.029 70.50% set_robust_list 1 0.008 0.008 0.008 0.008 0.00% madvise 1 0.005 0.005 0.005 0.005 0.00% clone 1 0.000 0.000 0.000 0.000 0.00% # Committer notes: Fix build on fedora:24-x-ARC-uClibc, debian:experimental-x-mips, debian:experimental-x-mipsel, ubuntu:16.04-x-arm and ubuntu:16.04-x-powerpc CC /tmp/build/perf/bench/epoll-ctl.o bench/epoll-ctl.c: In function 'init_fdmaps': bench/epoll-ctl.c:214:16: error: comparison between signed and unsigned integer expressions [-Werror=sign-compare] for (i = 0; i < nfds; i+=inc) { ^ bench/epoll-ctl.c: In function 'bench_epoll_ctl': bench/epoll-ctl.c:377:16: error: comparison between signed and unsigned integer expressions [-Werror=sign-compare] for (i = 0; i < nthreads; i++) { ^ bench/epoll-ctl.c:388:16: error: comparison between signed and unsigned integer expressions [-Werror=sign-compare] for (i = 0; i < nthreads; i++) { ^ cc1: all warnings being treated as errors Signed-off-by: Davidlohr Bueso Tested-by: Arnaldo Carvalho de Melo Cc: Andrew Morton Cc: Davidlohr Bueso Cc: Jason Baron Link: http://lkml.kernel.org/r/20181106152226.20883-3-dave@stgolabs.net [ Use inttypes.h to print rlim_t fields, fixing the build on Alpine Linux / musl libc ] [ Check if eventfd() is available, i.e. if HAVE_EVENTFD is defined ] Signed-off-by: Arnaldo Carvalho de Melo

perf bench: Add epoll parallel epoll_wait benchmark

2018-11-22T01:38:47Z

This program benchmarks concurrent epoll_wait(2) for file descriptors that are monitored with with EPOLLIN along various semantics, by a single epoll instance. Such conditions can be found when using single/combined or multiple queuing when load balancing. Each thread has a number of private, nonblocking file descriptors, referred to as fdmap. A writer thread will constantly be writing to the fdmaps of all threads, minimizing each threads's chances of epoll_wait not finding any ready read events and blocking as this is not what we want to stress. Full details in the start of the C file. Committer testing: # perf bench Usage: perf bench [] [] # List of all available benchmark collections: sched: Scheduler and IPC benchmarks mem: Memory access benchmarks numa: NUMA scheduling and MM benchmarks futex: Futex stressing benchmarks epoll: Epoll stressing benchmarks all: All benchmarks # perf bench epoll # List of available benchmarks for collection 'epoll': wait: Benchmark epoll concurrent epoll_waits all: Run all futex benchmarks # perf bench epoll wait # Running 'epoll/wait' benchmark: Run summary [PID 19295]: 3 threads monitoring on 64 file-descriptors for 8 secs. [thread 0] fdmap: 0xdaa650 ... 0xdaa74c [ 328241 ops/sec ] [thread 1] fdmap: 0xdaa900 ... 0xdaa9fc [ 351695 ops/sec ] [thread 2] fdmap: 0xdaabb0 ... 0xdaacac [ 381423 ops/sec ] Averaged 353786 operations/sec (+- 4.35%), total secs = 8 # Committer notes: Fix the build on debian:experimental-x-mips, debian:experimental-x-mipsel and others: CC /tmp/build/perf/bench/epoll-wait.o bench/epoll-wait.c: In function 'writerfn': bench/epoll-wait.c:399:12: error: format '%ld' expects argument of type 'long int', but argument 2 has type 'size_t' {aka 'unsigned int'} [-Werror=format=] printinfo("exiting writer-thread (total full-loops: %ld)\n", iter); ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~~~ bench/epoll-wait.c:86:31: note: in definition of macro 'printinfo' do { if (__verbose) { printf(fmt, ## arg); fflush(stdout); } } while (0) ^~~ cc1: all warnings being treated as errors Signed-off-by: Davidlohr Bueso Tested-by: Arnaldo Carvalho de Melo Cc: Andrew Morton Cc: Davidlohr Bueso Cc: Jason Baron Link: http://lkml.kernel.org/r/20181106152226.20883-2-dave@stgolabs.net Link: http://lkml.kernel.org/r/20181106182349.thdkpvshkna5vd7o@linux-r8p5> [ Applied above fixup as per Davidlohr's request ] [ Use inttypes.h to print rlim_t fields, fixing the build on Alpine Linux / musl libc ] [ Check if eventfd() is available, i.e. if HAVE_EVENTFD is defined ] Signed-off-by: Arnaldo Carvalho de Melo