diff --git a/README.md b/README.md index 30bb3ac..b4d9d1a 100644 --- a/README.md +++ b/README.md @@ -46,7 +46,7 @@ USAGE: pmcarch {-C CPU | -p PID | -c CMD} [interval [duration]] pmcarch -c 'cksum /boot/*' # measure run and measure this cmd -## All PMCs +## Extended PMCs These are available on some AWS Nitro hypervisor instance types. Eg, c5.9xl. @@ -91,3 +91,60 @@ USAGE: tlbstat {-C CPU | -p PID | -c CMD} [interval [duration]] tlbstat -p 181 # measure PID 181 only tlbstat -c 'cksum /boot/*' # measure run and measure this cmd + +## ALL PMCs + +These may not work on the cloud yet, and may be works-in-progress. They may also require more recent Linux versions for the PMC aliases. + +### cpucache + +Summarizes CPU L1/L2/LLC cache hit ratios. Eg: + +
+# ./cpucache +All counter columns are x 1000 +CYCLES INSTR IPC L1DREF L1DMISS L1D% L2REF L2MISS L2% LLCREF LLCMISS LLC% +55827883 44795258 0.80 11674473 98067 99.16 161026 78908 51.00 97522 31338 67.87 +55841905 44803990 0.80 11685292 98216 99.16 162127 79934 50.70 98810 30975 68.65 +55832518 44838601 0.80 11700599 98181 99.16 160586 78725 50.98 96895 31497 67.49 +55832741 44831064 0.80 11700963 98102 99.16 160849 78377 51.27 97296 31477 67.65 +55832729 44763451 0.80 11672814 98216 99.16 163742 80630 50.76 99961 30679 69.31 +55847526 44815072 0.80 11691652 98299 99.16 162176 79552 50.95 99092 30798 68.92 +55832810 44748131 0.80 11667573 98132 99.16 163209 82497 49.45 101771 31035 69.50 +55832325 44788718 0.80 11670262 98459 99.16 161834 79713 50.74 97504 31057 68.15 +55832453 44795688 0.80 11682541 98069 99.16 162384 77339 52.37 95222 30347 68.13 +55832224 44749672 0.80 11688124 98007 99.16 163076 79868 51.02 100960 30842 69.45 +55837742 44793089 0.80 11689272 98273 99.16 161173 79280 50.81 97667 31776 67.46 +[...] ++ +- 'CYCLES': CPU Cycles x 1000 +- 'INSTR': CPU Instructions x 1000 +- 'IPC': Instructions-Per-Cycle +- 'L1DREF': Level 1 data cache loads x 1000 +- 'L1DMISS': Level 1 data cache load misses x 1000 +- 'L1D%': Level 1 data cache hit ratio, as a percentage +- 'L2REF': Level 2 requests x 1000 +- 'L2MISS': Level 2 misses x 1000 +- 'L2%': Level 2 hit ratio, as a percentage +- 'LLCREF': Last Level Cache references x 1000 +- 'LLCMISS': Last Level Cache misses x 1000 +- 'LLC%': Last Level Cache hit ratio, as a percentage + +USAGE: + +
+# ./cpucache -h
+USAGE: cpucache {-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,
+ cpucache # show stats across all CPUs
+ cpucache 5 # show stats every 5 seconds
+ cpucache -C 0 # measure CPU 0 only
+ cpucache -p 181 # measure PID 181 only
+ cpucache -c 'cksum /boot/*' # measure run and measure this cmd
+
diff --git a/cpucache b/cpucache
new file mode 100755
index 0000000..c683fc7
--- /dev/null
+++ b/cpucache
@@ -0,0 +1,126 @@
+#!/bin/bash
+#
+# cpucache - measure and summarize CPU cache references (L1/L2/LLC).
+# Uses Linux perf and PMCs.
+#
+# FROM: https://github.com/brendangregg/pmc-cloud-tools
+#
+# USAGE: cpucache {-C CPU | -p PID | -c CMD} [interval [duration]]
+#
+# Columns:
+#
+# - CYCLES: CPU Cycles x 1000
+# - INSTR: CPU Instructions x 1000
+# - IPC: Instructions-Per-Cycle
+# - L1DREF: Level 1 data cache loads x 1000
+# - L1DMISS: Level 1 data cache load misses x 1000
+# - L1D%: Level 1 data cache hit ratio, as a percentage
+# - L2REF: Level 2 requests x 1000
+# - L2MISS: Level 2 misses x 1000
+# - L2%: Level 2 hit ratio, as a percentage
+# - LLCREF: Last Level Cache references x 1000
+# - LLCMISS: Last Level Cache misses x 1000
+# - LLC%: Last Level Cache hit ratio, as a percentage
+#
+# Copyright 2018 Netflix, Inc.
+# Licensed under the Apache License, Version 2.0 (the "License")
+#
+# 12-Jan-2018 Brendan Gregg Created this.
+
+function usage {
+ cat <<-END >&2
+ USAGE: cpucache {-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,
+ cpucache # show stats across all CPUs
+ cpucache 5 # show stats every 5 seconds
+ cpucache -C 0 # measure CPU 0 only
+ cpucache -p 181 # measure PID 181 only
+ cpucache -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
+# cycles are twice as a workaround for an issue
+# the r4f2e and r412e counters are from the architectural set, so should be stable
+echo "All counter columns are x 1000"
+perf stat -e cycles -e cycles \
+ -e L1-dcache-loads -e L1-dcache-load-misses \
+ -e l2_rqsts.references -e l2_rqsts.miss \
+ -e r4f2e -e r412e -e instructions \
+ -I $(( secs * 1000 )) $target 2>&1 | awk -v hlines=$hlines '
+ BEGIN {
+ htxt = sprintf("%-10s %-10s %5s %-9s %-9s %5s %-8s %-8s %5s %-8s %-8s %5s",
+ "CYCLES", "INSTR", "IPC", "L1DREF", "L1DMISS", "L1D%",
+ "L2REF", "L2MISS", "L2%", "LLCREF", "LLCMISS", "LLC%");
+ print htxt
+ header = hlines
+ }
+ /invalid/ { print $0 } # unsupported event
+ { gsub(/,/, ""); }
+ $3 == "cycles" { cycles = $2 }
+ $3 == "r4f2e" { llcref = $2 }
+ $3 == "r412e" { llcmiss = $2 }
+ $3 == "L1-dcache-loads" { l1dload = $2 }
+ $3 == "L1-dcache-load-misses" { l1dloadmiss = $2 }
+ $3 == "l2_rqsts.references" { l2ref = $2 }
+ $3 == "l2_rqsts.miss" { l2miss = $2 }
+ $3 == "instructions" {
+ if (--header == 0) {
+ print htxt
+ header = hlines
+ }
+
+ ins = $2
+ # if we are missing refs, then also reset hits to zero, to avoid a bogus ratio:
+ if (l1dload == 0) { l1dload = 1; l1dloadmiss = 0 }
+ if (l2ref == 0) { l2ref = 1; l2miss = 0 }
+ if (llcref == 0) { llcref = 1; llcmiss = 0 }
+ if (cycles == 0) { cycles = 1; ins = 0 } # PMCs are broken, or no events
+ l1dratio = 100 * (l1dload - l1dloadmiss) / l1dload
+ l2ratio = 100 * (l2ref - l2miss) / l2ref
+ llcratio = 100 * (llcref - llcmiss) / llcref
+
+ printf("%-10d %-10d %5.2f %-9d %-9d %5.2f %-8d %-8d %5.2f %-8d %-8d %5.2f\n",
+ cycles / 1000, ins / 1000, ins / cycles,
+ l1dload / 1000, l1dloadmiss / 1000, l1dratio,
+ l2ref / 1000, l2miss / 1000, l2ratio, llcref / 1000, llcmiss / 1000, llcratio)
+ }
+'