#!/bin/bash # # pmcipc - measure and summarize instructions per cycle. Uses Linux perf. # # This deliberately measures IPC only, for instances where measuring more than # a few PMCs leads to bogus measurements. Cross-check with pmcarch and # perf stat. # # FROM: https://github.com/brendangregg/pmc-cloud-tools # # USAGE: pmcipc {-C CPU | -p PID | -c CMD} [interval [duration]] # # PMCs are performance monitoring counters. This tool reports: # # - K_CYCLES: CPU Cycles x 1000 # - K_INSTR: CPU Instructions x 1000 # - IPC: Instructions-Per-Cycle # # Copyright 2019 Netflix, Inc. # Licensed under the Apache License, Version 2.0 (the "License") # # 04-Oct-2019 Brendan Gregg Created this. function usage { cat <<-END >&2 USAGE: pmcipc {-C CPU | -p PID | -c CMD} [interval [duration]] -C CPU # measure this CPU only -p PID # measure this PID only -c 'CMD' # measure this command only (quote it) interval # output interval in secs (default 1) duration # total seconds (default infinityish) eg, pmcipc # show stats across all CPUs pmcipc 5 # show stats every 5 seconds pmcipc -C 0 # measure CPU 0 only pmcipc -p 181 # measure PID 181 only pmcipc -c 'cksum /boot/*' # measure run and measure this cmd END exit } opt_cpu=0; opt_pid=0; opt_cmd=0; cpu=""; pid=""; cmd="" while getopts C:p:c:h opt do case $opt in C) opt_cpu=1; cpu=$OPTARG ;; p) opt_pid=1; pid=$OPTARG ;; c) opt_cmd=1; cmd=$OPTARG ;; h|?) usage ;; esac done shift $(( $OPTIND - 1 )) if (( opt_cpu + opt_pid + opt_cmd > 1 )); then echo >&2 "ERROR: pick one of -C, -p, -c" usage fi secs=${1:-1} # default 1 second duration=${2:-999999999} # default semi-infinite seconds hlines=25 # lines to repeat header target=-a (( opt_cpu )) && target="-C $cpu sleep $duration" (( opt_pid )) && target="-p $pid sleep $duration" (( opt_cmd )) && target="$cmd" if (( opt_pid )); then if [ ! -d /proc/$pid ]; then echo >&2 "ERROR: Can't find PID $pid. Exiting." exit fi fi # note that instructions is last on purpose, it triggers output # order of PMCs attempts to workaround a 3.13 issue. # cycles are twice as a workaround for an issue. # only three (<=4) PMCs are used to workaround another issue. # the raw counters are from the architectural set, so should be stable. perf stat -e cycles -e cycles -e instructions \ -I $(( secs * 1000 )) $target 2>&1 | awk -v hlines=$hlines ' BEGIN { htxt = sprintf("%-10s %-10s %5s", "K_CYCLES", "K_INSTR", "IPC"); print htxt header = hlines } /invalid/ { print $0 } # unsupported event { gsub(/,/, ""); } $3 == "cycles" { cycles = $2; } $3 == "instructions" { if (--header == 0) { print htxt header = hlines } ins = $2 if (llcref == 0) { llcref = 1; llcmiss = 0 } if (cycles == 0) { cycles = 1 } # PMCs are broken, or no events printf("%-10d %-10d %5.2f\n", cycles / 1000, ins / 1000, ins / cycles) } '