Re: [PATCH V1 2/4] selftests: amd-pstate: Trigger tbench benchmark and test cpus

From: Huang Rui
Date: Mon Jul 18 2022 - 20:15:41 EST


On Wed, Jul 06, 2022 at 03:36:20PM +0800, Meng, Li (Jassmine) wrote:
> Add tbench.sh trigger the tbench testing and monitor the cpu desire
> performance, frequency, load, power consumption and throughput etc.
>
> Signed-off-by: Meng Li <li.meng@xxxxxxx>
> ---
> tools/testing/selftests/amd-pstate/Makefile | 10 +-
> tools/testing/selftests/amd-pstate/main.sh | 168 ++++++++++++++++-
> tools/testing/selftests/amd-pstate/tbench.sh | 187 +++++++++++++++++++
> 3 files changed, 361 insertions(+), 4 deletions(-)
> create mode 100755 tools/testing/selftests/amd-pstate/tbench.sh
>
> diff --git a/tools/testing/selftests/amd-pstate/Makefile b/tools/testing/selftests/amd-pstate/Makefile
> index 167ab51ec290..e2fb03381a32 100644
> --- a/tools/testing/selftests/amd-pstate/Makefile
> +++ b/tools/testing/selftests/amd-pstate/Makefile
> @@ -4,7 +4,15 @@
> # No binaries, but make sure arg-less "make" doesn't trigger "run_tests"
> all:
>
> +uname_M := $(shell uname -m 2>/dev/null || echo not)
> +ARCH ?= $(shell echo $(uname_M) | sed -e s/i.86/x86/ -e s/x86_64/x86/)
> +
> +ifeq (x86,$(ARCH))
> +TEST_GEN_FILES += ../../../power/x86/amd_pstate_tracer/amd_pstate_trace.py
> +TEST_GEN_FILES += ../../../power/x86/intel_pstate_tracer/intel_pstate_tracer.py
> +endif
> +
> TEST_PROGS := run.sh
> -TEST_FILES := main.sh basic.sh
> +TEST_FILES := main.sh basic.sh tbench.sh
>
> include ../lib.mk
> diff --git a/tools/testing/selftests/amd-pstate/main.sh b/tools/testing/selftests/amd-pstate/main.sh
> index 1c28b5d7b4c5..ae71fe62e3b0 100755
> --- a/tools/testing/selftests/amd-pstate/main.sh
> +++ b/tools/testing/selftests/amd-pstate/main.sh
> @@ -9,6 +9,7 @@ else
> fi
>
> source basic.sh
> +source tbench.sh
>
> # amd-pstate-ut only run on x86/x86_64 AMD systems.
> ARCH=$(uname -m 2>/dev/null | sed -e 's/i.86/x86/' -e 's/x86_64/x86/')
> @@ -16,10 +17,95 @@ VENDOR=$(cat /proc/cpuinfo | grep -m 1 'vendor_id' | awk '{print $NF}')
>
> FUNC=all
> OUTFILE=selftest
> +OUTFILE_TBENCH="$OUTFILE.tbench"
> +
> +SYSFS=
> +CPUROOT=
> +CPUFREQROOT=
> +MAKE_CPUS=
> +
> +TIME_LIMIT=100
> +PROCESS_NUM=128
> +LOOP_TIMES=3
> +TRACER_INTERVAL=10
>
> # Kselftest framework requirement - SKIP code is 4.
> ksft_skip=4
>
> +# Counts CPUs with cpufreq directories
> +count_cpus()
> +{
> + count=0;
> +
> + for cpu in `ls $CPUROOT | grep "cpu[0-9].*"`; do
> + if [ -d $CPUROOT/$cpu/cpufreq ]; then
> + let count=count+1;
> + fi
> + done
> +
> + echo $count;
> +}
> +
> +# $1: policy
> +find_current_governor()
> +{
> + cat $CPUFREQROOT/$1/scaling_governor
> +}
> +
> +# $1: policy
> +backup_governor()
> +{
> + local cur_gov=$(find_current_governor $1)
> +
> + printf "Governor backup done for $1: $cur_gov\n"
> + echo "$1 $cur_gov" >> $OUTFILE.backup_governor.log
> +}
> +
> +restore_governor()
> +{
> + i=0;
> +
> + policies=$(awk '{print $1}' $OUTFILE.backup_governor.log)
> + for policy in $policies; do
> + let i++;
> + governor=$(sed -n ''$i'p' $OUTFILE.backup_governor.log | awk '{print $2}')
> +
> + # switch governor
> + _switch_governor $policy $governor
> + printf "Governor restored for $policy to $governor\n"
> + done
> +}
> +
> +# $1: policy, $2: governor
> +_switch_governor()
> +{
> + echo $2 > $CPUFREQROOT/$1/scaling_governor
> +}
> +
> +for_each_policy()
> +{
> + policies=$(ls $CPUFREQROOT| grep "policy[0-9].*")
> + for policy in $policies; do
> + $@ $policy
> + done
> +}
> +
> +# $1: governor, $2: policy
> +switch_governor()
> +{
> + local filepath=$CPUFREQROOT/$2/scaling_available_governors
> +
> + # Exit if cpu isn't managed by cpufreq core
> + if [ ! -f $filepath ]; then
> + return;
> + fi
> +
> + # switch governor
> + _switch_governor $2 $1
> +
> + printf "Switched governor for $2 to $1\n"
> +}
> +
> # All amd-pstate tests
> amd_pstate_all()
> {
> @@ -27,8 +113,19 @@ amd_pstate_all()
> printf "***** Running AMD P-state Sanity Tests *****\n"
> printf "=============================================\n\n"
>
> + count=$(count_cpus)
> + if [ $count = 0 ]; then
> + printf "No cpu is managed by cpufreq core, exiting\n"
> + exit;
> + else
> + printf "AMD P-state manages: $count CPUs\n"
> + fi
> +
> # unit test for amd-pstate kernel driver
> amd_pstate_basic
> +
> + # tbench
> + amd_pstate_tbench
> }
>
> helpme()
> @@ -37,21 +134,26 @@ helpme()
> [-h <help>]
> [-o <output-file-for-dump>]
> [-c <all: All testing,
> - basic: Basic testing.>]
> + basic: Basic testing,
> + tbench: Tbench testing.>]
> + [-t <tbench time limit>]
> + [-p <tbench process number>]
> + [-l <loop times for tbench]
> + [-i <amd tracer interval]
> \n"
> exit 2
> }
>
> parse_arguments()
> {
> - while getopts ho:c: arg
> + while getopts ho:c:t:p:l:i: arg
> do
> case $arg in
> h) # --help
> helpme
> ;;
>
> - c) # --func_type (Function to perform: basic (default: all))
> + c) # --func_type (Function to perform: basic, tbench (default: all))
> FUNC=$OPTARG
> ;;
>
> @@ -59,6 +161,21 @@ parse_arguments()
> OUTFILE=$OPTARG
> ;;
>
> + t) # --tbench-time-limit
> + TIME_LIMIT=$OPTARG
> + ;;
> +
> + p) # --tbench-process-number
> + PROCESS_NUM=$OPTARG
> + ;;
> +
> + l) # --tbench-loop-times
> + LOOP_TIMES=$OPTARG
> + ;;
> +
> + i) # --amd-tracer-interval
> + TRACER_INTERVAL=$OPTARG
> + ;;
> \?)
> helpme
> ;;
> @@ -91,10 +208,44 @@ prerequisite()
> echo $msg please run this as root >&2
> exit $ksft_skip
> fi
> +
> + if [ ! -f perf ] ; then
> + echo $msg please build perf under directory tools/perf and copy it to directory amd-pstate. >&2
> + exit $ksft_skip
> + fi
> +
> + SYSFS=`mount -t sysfs | head -1 | awk '{ print $3 }'`
> +
> + if [ ! -d "$SYSFS" ]; then
> + echo $msg sysfs is not mounted >&2
> + exit 2
> + fi
> +
> + CPUROOT=$SYSFS/devices/system/cpu
> + CPUFREQROOT="$CPUROOT/cpufreq"
> +
> + if ! ls $CPUROOT/cpu* > /dev/null 2>&1; then
> + echo $msg cpus not available in sysfs >&2
> + exit 2
> + fi
> +
> + if ! ls $CPUROOT/cpufreq > /dev/null 2>&1; then
> + echo $msg cpufreq directory not available in sysfs >&2
> + exit 2
> + fi
> }
>
> do_test()
> {
> + # Check if CPUs are managed by cpufreq or not
> + count=$(count_cpus)
> + MAKE_CPUS=$((count*2))
> +
> + if [ $count = 0 ]; then
> + echo "No cpu is managed by cpufreq core, exiting"
> + exit 2;
> + fi
> +
> case "$FUNC" in
> "all")
> amd_pstate_all
> @@ -104,6 +255,10 @@ do_test()
> amd_pstate_basic
> ;;
>
> + "tbench")
> + amd_pstate_tbench
> + ;;
> +
> *)
> echo "Invalid [-f] function type"
> helpme
> @@ -122,10 +277,17 @@ pre_clear_dumps()
> "basic")
> ;;
>
> + "tbench")
> + rm -rf $OUTFILE.log
> + rm -rf $OUTFILE.backup_governor.log
> + rm -rf $OUTFILE_TBENCH*
> + ;;
> +
> esac
> }
>
> post_clear_dumps()
> {
> rm -rf $OUTFILE.log
> + rm -rf $OUTFILE.backup_governor.log
> }
> diff --git a/tools/testing/selftests/amd-pstate/tbench.sh b/tools/testing/selftests/amd-pstate/tbench.sh
> new file mode 100755
> index 000000000000..0320300971ce
> --- /dev/null
> +++ b/tools/testing/selftests/amd-pstate/tbench.sh
> @@ -0,0 +1,187 @@
> +#!/bin/sh
> +# SPDX-License-Identifier: GPL-2.0
> +
> +# Testing and monitor the cpu frequency and performance when
> +# this script trigger tbench test.
> +
> +# protect against multiple inclusion
> +if [ $FILE_TBENCH ]; then
> + return 0
> +else
> + FILE_TBENCH=DONE
> +fi
> +
> +tbench_name="dbench-4.0"
> +tbench_tar="$tbench_name.tar.gz"
> +tbench_url="https://ftp.samba.org/pub/pub/tridge/dbench/dbench-4.0.tar.gz";
> +tbench_governors=("ondemand" "schedutil")
> +
> +# $1: governor, $2: round, $3: des-perf, $4: freq, $5: load, $6: performance, $7: energy, $8: PPW
> +store_csv_tbench()
> +{
> + echo "$1, $2, $3, $4, $5, $6, $7, $8" | tee -a $OUTFILE_TBENCH.csv > /dev/null 2>&1
> +}
> +
> +empty_line_csv_tbench()
> +{
> + echo "" | tee -a $OUTFILE_TBENCH.csv > /dev/null 2>&1
> +}
> +
> +pre_clear_tbench()
> +{
> + rm -rf results/tracer-tbench*
> +}
> +
> +post_clear_tbench()
> +{
> + rm -rf results/tracer-tbench*
> + rm -rf $OUTFILE_TBENCH*.log
> +}
> +
> +install_tbench()
> +{
> + if [ ! -d $tbench_name ]; then
> + printf "Download tbench, please wait a moment ...\n\n"
> + wget -O $tbench_tar $tbench_url > /dev/null 2>&1
> +
> + printf "Tar tbench ...\n\n"
> + tar -xzf $tbench_tar
> +
> + printf "Compile and install tbench ...\n\n"
> + cd $tbench_name
> + ./autogen.sh
> + ./configure > /dev/null 2>&1
> + make > /dev/null 2>&1
> + make install > /dev/null 2>&1
> + cd ..
> + fi
> +}
> +
> +# $1: governor, $2: loop
> +run_tbench()
> +{
> + echo "Launching amd pstate tracer for $1 #$2 tracer_interval: $TRACER_INTERVAL"
> + ./amd_pstate_trace.py -n tracer-tbench-$1-$2 -i $TRACER_INTERVAL > /dev/null 2>&1 &
> +
> + printf "Test tbench for $1 #$2 time_limit: $TIME_LIMIT procs_num: $PROCESS_NUM\n"
> + cd $tbench_name
> + ./tbench_srv > /dev/null 2>&1 &
> + ./../perf stat -a --per-socket -I 1000 -e power/energy-pkg/ ./tbench -t $TIME_LIMIT $PROCESS_NUM > ../$OUTFILE_TBENCH-perf-$1-$2.log 2>&1
> + cd ..
> +
> + pid=`pidof tbench_srv`
> + kill $pid
> +
> + for job in `jobs -p`
> + do
> + echo "Waiting for job id $job"
> + wait $job
> + done
> +}
> +
> +# $1: governor, $2: loop
> +parse_tbench()
> +{
> + awk '{print $5}' results/tracer-tbench-$1-$2/cpu.csv | sed -e '1d' | sed s/,// > $OUTFILE_TBENCH-des-perf-$1-$2.log
> + avg_des_perf=$(awk 'BEGIN {i=0; sum=0};{i++; sum += $1};END {print sum/i}' $OUTFILE_TBENCH-des-perf-$1-$2.log)
> + printf "Tbench-$1-#$2 avg des perf: $avg_des_perf\n" | tee -a $OUTFILE_TBENCH.result
> +
> + awk '{print $7}' results/tracer-tbench-$1-$2/cpu.csv | sed -e '1d' | sed s/,// > $OUTFILE_TBENCH-freq-$1-$2.log
> + avg_freq=$(awk 'BEGIN {i=0; sum=0};{i++; sum += $1};END {print sum/i}' $OUTFILE_TBENCH-freq-$1-$2.log)
> + printf "Tbench-$1-#$2 avg freq: $avg_freq\n" | tee -a $OUTFILE_TBENCH.result
> +
> + awk '{print $11}' results/tracer-tbench-$1-$2/cpu.csv | sed -e '1d' | sed s/,// > $OUTFILE_TBENCH-load-$1-$2.log
> + avg_load=$(awk 'BEGIN {i=0; sum=0};{i++; sum += $1};END {print sum/i}' $OUTFILE_TBENCH-load-$1-$2.log)
> + printf "Tbench-$1-#$2 avg load: $avg_load\n" | tee -a $OUTFILE_TBENCH.result
> +
> + grep Throughput $OUTFILE_TBENCH-perf-$1-$2.log | awk '{print $2}' > $OUTFILE_TBENCH-throughput-$1-$2.log
> + tp_sum=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum}' $OUTFILE_TBENCH-throughput-$1-$2.log)
> + printf "Tbench-$1-#$2 throughput(MB/s): $tp_sum\n" | tee -a $OUTFILE_TBENCH.result
> +
> + grep Joules $OUTFILE_TBENCH-perf-$1-$2.log | awk '{print $4}' > $OUTFILE_TBENCH-energy-$1-$2.log
> + en_sum=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum}' $OUTFILE_TBENCH-energy-$1-$2.log)
> + printf "Tbench-$1-#$2 power consumption(J): $en_sum\n" | tee -a $OUTFILE_TBENCH.result
> +
> + ppw=`echo "scale=4;($TIME_LIMIT-1)*$tp_sum/$en_sum" | bc | awk '{printf "%.4f", $0}'`
> + printf "Tbench-$1-#$2 PPW(MB/J): $ppw\n" | tee -a $OUTFILE_TBENCH.result
> + printf "\n" | tee -a $OUTFILE_TBENCH.result

Please clarify PPW = performance per watt here.

> +
> + store_csv_tbench $1 $2 $avg_des_perf $avg_freq $avg_load $tp_sum $en_sum $ppw
> +}
> +
> +# $1: governor
> +loop_tbench()
> +{
> + printf "\nTbench total test times is $LOOP_TIMES for $1\n\n"
> + for i in `seq 1 $LOOP_TIMES`
> + do
> + run_tbench $1 $i
> + parse_tbench $1 $i
> + done
> +}
> +
> +# $1: governor
> +gather_tbench()
> +{
> + printf "Tbench test result for $1 (loops:$LOOP_TIMES)" | tee -a $OUTFILE_TBENCH.result
> + printf "\n--------------------------------------------------\n" | tee -a $OUTFILE_TBENCH.result
> +
> + grep "Tbench-$1-#" $OUTFILE_TBENCH.result | grep "avg des perf:" | awk '{print $NF}' > $OUTFILE_TBENCH-des-perf-$1.log
> + avg_des_perf=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum/'$LOOP_TIMES'}' $OUTFILE_TBENCH-des-perf-$1.log)
> + printf "Tbench-$1 avg des perf: $avg_des_perf\n" | tee -a $OUTFILE_TBENCH.result
> +
> + grep "Tbench-$1-#" $OUTFILE_TBENCH.result | grep "avg freq:" | awk '{print $NF}' > $OUTFILE_TBENCH-freq-$1.log
> + avg_freq=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum/'$LOOP_TIMES'}' $OUTFILE_TBENCH-freq-$1.log)
> + printf "Tbench-$1 avg freq: $avg_freq\n" | tee -a $OUTFILE_TBENCH.result
> +
> + grep "Tbench-$1-#" $OUTFILE_TBENCH.result | grep "avg load:" | awk '{print $NF}' > $OUTFILE_TBENCH-load-$1.log
> + avg_load=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum/'$LOOP_TIMES'}' $OUTFILE_TBENCH-load-$1.log)
> + printf "Tbench-$1 avg load: $avg_load\n" | tee -a $OUTFILE_TBENCH.result
> +
> + grep "Tbench-$1-#" $OUTFILE_TBENCH.result | grep "throughput(MB/s):" | awk '{print $NF}' > $OUTFILE_TBENCH-throughput-$1.log
> + tp_sum=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum}' $OUTFILE_TBENCH-throughput-$1.log)
> + printf "Tbench-$1 total throughput(MB/s): $tp_sum\n" | tee -a $OUTFILE_TBENCH.result
> +
> + avg_tp=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum/'$LOOP_TIMES'}' $OUTFILE_TBENCH-throughput-$1.log)
> + printf "Tbench-$1 avg throughput(MB/s): $avg_tp\n" | tee -a $OUTFILE_TBENCH.result
> +
> + grep "Tbench-$1-#" $OUTFILE_TBENCH.result | grep "power consumption(J):" | awk '{print $NF}' > $OUTFILE_TBENCH-energy-$1.log
> + en_sum=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum}' $OUTFILE_TBENCH-energy-$1.log)
> + printf "Tbench-$1 total power consumption(J): $en_sum\n" | tee -a $OUTFILE_TBENCH.result
> +
> + avg_en=$(awk 'BEGIN {sum=0};{sum += $1};END {print sum/'$LOOP_TIMES'}' $OUTFILE_TBENCH-energy-$1.log)
> + printf "Tbench-$1 avg power consumption(J): $avg_en\n" | tee -a $OUTFILE_TBENCH.result
> +
> + ppw=`echo "scale=4;($TIME_LIMIT-1)*$avg_tp/$avg_en" | bc | awk '{printf "%.4f", $0}'`
> + printf "Tbench-$1 PPW(MB/J): $ppw\n" | tee -a $OUTFILE_TBENCH.result
> + printf "\n" | tee -a $OUTFILE_TBENCH.result

It's better to add commment here to describe the formula that how to caculate the performance per watt (PPW) result for the tbench tests.

> +
> + store_csv_tbench $1 "Average" $avg_des_perf $avg_freq $avg_load $avg_tp $avg_en $ppw
> + empty_line_csv_tbench
> +}
> +
> +amd_pstate_tbench()
> +{
> + printf "\n---------------------------------------------\n"
> + printf "*** Running AMD P-state tbench ***"
> + printf "\n---------------------------------------------\n"
> +
> + pre_clear_tbench
> +
> + install_tbench
> +
> + store_csv_tbench "Governor" "Round" "Des-perf" "Freq" "Load" "Performance(MB/s)" "Energy(Joules)" "PPW(MB/J)"

The same comment here above, we would better to have abbreviation
desciption for PPW in the result.

Thanks,
Ray

> +
> + for_each_policy backup_governor
> +
> + for governor in ${tbench_governors[*]} ; do
> + printf "\nSpecified governor is $governor\n\n"
> + for_each_policy switch_governor $governor
> + loop_tbench $governor
> + gather_tbench $governor
> + done
> +
> + restore_governor
> +
> + post_clear_tbench
> +}
> --
> 2.25.1
>