cacheperf

#!/bin/sh

readonly COUNTERS="`echo "
	cpu-cycles:cpu-cycles
	instructions:instructions
	cpu/event=0x0e,umask=0x01,cmask=0x01,inv,any/:stalled-cycles (UOPS_ISSUED.ANY)
	L1-dcache-load-misses:L1-dcache-load-misses (L1D.REPLACEMENT)
	L1-icache-load-misses:L1-icache-load-misses (ICACHE.MISSES)
	dTLB-load-misses:dTLB-load-misses (DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK)
	iTLB-load-misses:iTLB-load-misses (ITLB_MISSES.MISS_CAUSES_A_WALK)
	r81d0:MEM_UOPS_RETIRED.ALL_LOADS (i.e. L1-dcache-loads, dTLB-load-misses)
	r08d1:MEM_LOAD_UOPS_RETIRED.L1_MISS
	r10d1:MEM_LOAD_UOPS_RETIRED.L2_MISS
	r20d1:MEM_LOAD_UOPS_RETIRED.L3_MISS (BDE70)
	r2124:L2_RQSTS.DEMAND_DATA_RD_MISS
	re124:L2_RQSTS.ALL_DEMAND_DATA_RD
	re424:L2_RQSTS.ALL_CODE_RD
	r01f0:L2_TRANS.DEMAND_DATA_RD
	r04f0:L2_TRANS.CODE_RD
	r2424:L2_RQSTS.CODE_RD_MISS (HSW)
	r2724:L2_RQSTS.ALL_DEMAND_MISS (HSW)
	re724:L2_RQSTS.ALL_DEMAND_REFERENCES (HSW)
	r3f24:L2_RQSTS.MISS (HSW)
	rff24:L2_RQSTS.REFERENCES (HSW)
	r4f2e:LLC_REFERENCE (BDE70)
	r412e:LLC_MISS (BDE70)
	" | tail -n+2 | tr -d '\t'`"

if [ $# -ne 2 -a \( $# -ne 1 -o "$1" != "end" \) ]
then
	cat <<-tac
		USAGE: $0 <-r|-s> <<list,of,cores> | end>

			-r means record a perf.data file
			-s means only output aggregate statistics

		Records cache-related performance counters on the specified cores
		until called again with the 'end' argument.
		Multiple invocations "stack."
	tac
	exit 1
fi

cores="$2"
case "$1" in
-r)
	mode="record"
	;;
-s)
	mode="stat"
	;;
end)
	cores="$1"
	;;
*)
	printf "%s\n" "$1: Unknown switch (expected -r or -s)" >&2
	exit 1
	;;
esac
set -e

launch_record() {
	rm -f "perf.data.$cores"
	ln -sf "perf.data.$cores" perf.data
	sudo perf record -se"`echo "$COUNTERS" | cut -d: -f1 | paste -sd,`" --call-graph lbr -aC"$cores" -o perf.data time -f" %e seconds time elapsed" sleep inf 2>&1 &
}

report_record() {
	perf report --stdio -iperf.data | grep -A1 '^# Samples:' | grep -v -- -- | sed "s/'//g" | rev | cut -d" " -f1 | rev | paste - - | eval sed `echo "$COUNTERS" | sed 's|.*|-e"s:&:"|'`
}

launch_stat() {
	sudo perf stat -e"`echo "$COUNTERS" | cut -d: -f1 | paste -sd,`" 2>&1 -aC"$cores" | tee "perf.dat.$cores" | tail -n2 | tr -s " " &
}

report_stat() {
	eval sed `echo "$COUNTERS" | sed 's|.*|-e"s:&:"|'` perf.dat | tail -n+4 | head -n-3 | tr -s " " | sed 's/ \([^ ]\+\) \([^#]\+\) .*/\2\t\1/' | tr -d ,
}

findproc() {
	local regex="$1"
	set -e

	ps -eocomm,pgid,pid,cmd | tr -s " " | grep "$regex" | tail -n1
}

if [ "$cores" = "end" ]
then
	perfline="`findproc "^perf "`"
	if [ -z "$perfline" ]
	then
		echo "No instances left to end"
		exit 0
	fi

	perfpgid="`printf %s "$perfline" | cut -d" " -f2`"
	perfpid="`printf %s "$perfline" | cut -d" " -f3`"
	chldpid="`findproc "^sleep $perfpgid .*inf$" | cut -d" " -f3`"
	if [ -n "$chldpid" ]
	then
		targetpid="$chldpid"
	else
		targetpid="$perfpid"
	fi

	sudo kill -int "$targetpid"
	while ps -eopid | tr -d " " | grep -x "$perfpid" >/dev/null; do :; done

	mode="`printf %s "$perfline" | sed 's/.\+perf \([^ ]\+\).*/\1/'`"
	cores="`printf %s "$perfline" | sed 's/.*-aC\([^ ]\+\).*/\1/'`"
	case "$mode" in
	record)
		ln -sf "perf.data.$cores" perf.data
		sudo chown "`whoami`:" perf.data
		mode="record"
		;;
	stat)
		ln -sf "perf.dat.$cores" perf.dat
		mode="stat"
		;;
	*)
		echo "INTERNAL ERROR: Neither record nor stat?!" >&2
		exit 2
	esac
	"report_$mode"
else
	"launch_$mode"
fi