Re: [Sched-tools] [RESEND PATCH v2] doc: measure the efficiency of cpufreq governors

Hi Xunlei,

On 18 June 2015 at 11:06, Xunlei Pang xlpang@126.com wrote:

From: Xunlei Pang pang.xunlei@linaro.org

DVFS adds a latency in the execution of task because of the time to decide to move at max freq. We need to measure this latency and check that the governor stays in an acceptable range.

When workgen runs a json file, a log file is created for each thread.

you use rt-app and not workgen in your script

This log file records the number of loop that has been executed and the duration for executing these loops (per phase). We can use these figures to evaluate to latency that is added by a cpufreq governor and its "performance efficiency".

We use the run+sleep pattern to do the measurement, for the run time per loop, the performance governor should run the expected duration as the CPU stays a max freq. At the opposite, the powersave governor will give use the longest duration (as it stays at lowest OPP). Other governor will be somewhere between the 2 previous duration as they will use several OPP and will go back to max frequency after a defined duration which depends on its monitoring period.

The formula:

duration of powersave gov - duration of the gov

-------------------------------------------------------- x 100% duration of powersave gov - duration of performance gov

will give the efficiency of the governor. 100% means as efficient as the perf governor and 0% means as efficient as the powersave governor.

This patch offers json files and shell scripts to do the measurement,

Usage: ./test.sh <cpus> <runtime> <sleeptime> cpus: number of cpus in the CPU0's frequency domain

Why do you need the number of cpus in CPU0's freq domain as an argument of the script ? The cpus that share the same freq domain, are available in /sys/devices/system/cpu/cpu0/cpufreq/related_cpus and/or affected_cpus. I don't remember the exact difference between both lists

And why do you need this parameter at all ? In your script, you set the governor of all cpus of the freq domain but AFAICT, cpus in the same freq domain, share the same governor so all cpus will use the new governor as soon as you change the governor of one cpu.

Then, how can i test the efficiency of a cpu that doesn't belong to CPU0 freq domain ? IMHO, you should better remove this parameter and add a new one to select the cpu on which you want to run the bench. Futhermore, CPU0 is the cpu that is the more used by default in a platform so it's worth to use another one in order to not be disturbed by background activity.

runtime: running time in ms per loop of the workload pattern sleeptime: sleeping time in ms per loop of the workload pattern

Example: "./test.sh 4 100 1000" means CPU0~CPU3 sharing frequency, "100ms run + 1000ms sleep" workload pattern.

test result on my machine: ~#./test.sh 4 100 1000 Frequency domain CPU0~CPU3, run 100ms, sleep 1000ms: powersave efficiency: 0% performance efficiency: 100% conservative efficiency: 28% ondemand efficiency: 95%

NOTE: Make sure there are "sed", "cut", "grep", "rt-app", etc tools on your test machine, and run the script under root privilege.

Signed-off-by: Xunlei Pang pang.xunlei@linaro.org

doc/examples/cpufreq_governor_efficiency/README | 54 ++++++++++++++ .../cpufreq_governor_efficiency/calibration.json | 27 +++++++ .../cpufreq_governor_efficiency/calibration.sh | 11 +++ doc/examples/cpufreq_governor_efficiency/dvfs.json | 27 +++++++ doc/examples/cpufreq_governor_efficiency/dvfs.sh | 38 ++++++++++ doc/examples/cpufreq_governor_efficiency/test.sh | 82 ++++++++++++++++++++++ 6 files changed, 239 insertions(+) create mode 100644 doc/examples/cpufreq_governor_efficiency/README create mode 100644 doc/examples/cpufreq_governor_efficiency/calibration.json create mode 100755 doc/examples/cpufreq_governor_efficiency/calibration.sh create mode 100644 doc/examples/cpufreq_governor_efficiency/dvfs.json create mode 100755 doc/examples/cpufreq_governor_efficiency/dvfs.sh create mode 100755 doc/examples/cpufreq_governor_efficiency/test.sh

diff --git a/doc/examples/cpufreq_governor_efficiency/README b/doc/examples/cpufreq_governor_efficiency/README new file mode 100644 index 0000000..cc8efe1 --- /dev/null +++ b/doc/examples/cpufreq_governor_efficiency/README @@ -0,0 +1,54 @@ +Measure the efficiency of cpufreq governors using rt-app

+BACKGROUND:

• DVFS adds a latency in the execution of task because of the time to
• decide to move at max freq. We need to measure this latency and check
• that the governor stays in an acceptable range.
• When workgen runs a json file, a log file is created for each thread.
• This log file records the number of loop that has been executed and
• the duration for executing these loops (per phase). We can use these
• figures to evaluate to latency that is added by a cpufreq governor
• and its "performance efficiency".
• We use the run+sleep pattern to do the measurement, for the run time per
• loop, the performance governor should run the expected duration as the
• CPU stays a max freq. At the opposite, the powersave governor will give
• use the longest duration (as it stays at lowest OPP). Other governor will
• be somewhere between the 2 previous duration as they will use several OPP
• and will go back to max frequency after a defined duration which depends
• on its monitoring period.
• The formula:

•    duration of powersave gov - duration of the gov
  

• -------------------------------------------------------- x 100%

• duration of powersave gov - duration of performance gov
  

• will give the efficiency of the governor. 100% means as efficient as
• the perf governor and 0% means as efficient as the powersave governor.
• This test offers json files and shell scripts to do the measurement,

+USAGE:

• ./test.sh <cpus> <runtime> <sleeptime>
• cpus: number of cpus in the CPU0's frequency domain
• runtime: running time in ms per loop of the workload pattern
• sleeptime: sleeping time in ms per loop of the workload pattern

+Example:

• "./test.sh 4 100 1000" means
• CPU0~CPU3 sharing frequency, "100ms run + 1000ms sleep" workload pattern.
• test result on an Intel machine:
• ~#./test.sh 4 100 1000
• Frequency domain CPU0~CPU3, run 100ms, sleep 1000ms:
• powersave efficiency: 0%
• performance efficiency: 100%
• conservative efficiency: 28%
• ondemand efficiency: 95%

+NOTE:

• Make sure there are "sed", "cut", "grep", "rt-app", etc tools on your test
• machine, and run the script under root privilege.

diff --git a/doc/examples/cpufreq_governor_efficiency/calibration.json b/doc/examples/cpufreq_governor_efficiency/calibration.json new file mode 100644 index 0000000..4377990 --- /dev/null +++ b/doc/examples/cpufreq_governor_efficiency/calibration.json @@ -0,0 +1,27 @@ +{

•   "tasks" : {
  

•           "thread" : {
  

•                   "instance" : 1,
  

•                   "cpus" : [0],
  

•                   "loop" : 1,
  

•                   "phases" : {
  

•                           "run" : {
  

•                                   "loop" : 1,
  

•                                   "run" : 200000,
  

•                           },
  

•                           "sleep" : {
  

•                                   "loop" : 1,
  

•                                   "sleep" : 200000,
  

•                           }
  

•                   }
  

•           }
  

•   },
  

•   "global" : {
  

•           "default_policy" : "SCHED_FIFO",
  

•           "calibration" : "CPU0",
  

•           "lock_pages" : true,
  

•           "ftrace" : true,
  

remove or disable the ftrace parameter. Otherwise rt-app will return an error if ftrace is not enable in the kernel. Then, your script stops without any message if rt-app fails to run the use case, you should detect the error and display a warning

•           "logdir" : "./",
  

•   }
  

+}

diff --git a/doc/examples/cpufreq_governor_efficiency/calibration.sh b/doc/examples/cpufreq_governor_efficiency/calibration.sh new file mode 100755 index 0000000..d10e644 --- /dev/null +++ b/doc/examples/cpufreq_governor_efficiency/calibration.sh @@ -0,0 +1,11 @@ +#!/bin/sh

+set -e

+echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor

+sleep 1

+pLoad=$(rt-app calibration.json 2>&1 |grep pLoad |sed 's/.*= (.*)ns.*/\1/') +sed 's/"calibration" : .*,/"calibration" : '$pLoad',/' -i dvfs.json

diff --git a/doc/examples/cpufreq_governor_efficiency/dvfs.json b/doc/examples/cpufreq_governor_efficiency/dvfs.json new file mode 100644 index 0000000..b413156 --- /dev/null +++ b/doc/examples/cpufreq_governor_efficiency/dvfs.json @@ -0,0 +1,27 @@ +{

•   "tasks" : {
  

•           "thread" : {
  

•                   "instance" : 1,
  

•                   "cpus" : [0],
  

•                   "loop" : 5,
  

i wonder if 5 loops are enough to let the system stabilize ? May be we can extract more statistic like the min/max/average/stdev value ? While testing your script, i have seen some variation on the results (especially with short run time)

•                   "phases" : {
  

•                           "running" : {
  

•                                   "loop" : 1,
  

•                                   "run" : 100000,
  

•                           },
  

•                           "sleeping" : {
  

•                                   "loop" : 1,
  

•                                   "sleep" : 1000000,
  

•                           }
  

•                   }
  

•           }
  

•   },
  

•   "global" : {
  

•           "default_policy" : "SCHED_OTHER",
  

•           "calibration" : 90,
  

•           "lock_pages" : true,
  

•           "ftrace" : true,
  

Remove or disable ftrace parameter

•           "logdir" : "./",
  

•   }
  

+}

diff --git a/doc/examples/cpufreq_governor_efficiency/dvfs.sh b/doc/examples/cpufreq_governor_efficiency/dvfs.sh new file mode 100755 index 0000000..8591fc7 --- /dev/null +++ b/doc/examples/cpufreq_governor_efficiency/dvfs.sh @@ -0,0 +1,38 @@ +#!/bin/sh

+#echo $1 $2 $3 +set -e

+if [ $1 ] && [ $2 ] ; then

•   for i in $(seq 0 1 $(expr $2 - 1)); do
  

•           echo $1 > /sys/devices/system/cpu/cpu$i/cpufreq/scaling_governor
  

•           #cat /sys/devices/system/cpu/cpu$i/cpufreq/scaling_governor
  

is it for debug purpose ?

•   done
  

•   sleep 3
  

+fi

+if [ $3 ] ; then

•   sed 's/"run" : .*,/"run" : '$3',/' -i dvfs.json
  

+fi

+if [ $4 ] ; then

•   sed 's/"sleep" : .*,/"sleep" : '$4',/' -i dvfs.json
  

+fi

+#cat dvfs.json

+rt-app dvfs.json 2> /dev/null

+if [ $1 ] ; then

•   mv -f rt-app-thread-0.log rt-app_$1_run$3us_sleep$4us.log
  

•   sum=0
  

•   for i in $(cat rt-app_$1_run$3us_sleep$4us.log | sed 'n;d' | sed '1d' |cut -f 3); do
  

•           sum=$(expr $sum + $i)
  

•   done
  

•   sum=$(expr $sum / 5)
  

•   echo $sum
  

•   rm -f rt-app_$1_run$3us_sleep$4us.log
  

+fi

diff --git a/doc/examples/cpufreq_governor_efficiency/test.sh b/doc/examples/cpufreq_governor_efficiency/test.sh new file mode 100755 index 0000000..d72fc6a --- /dev/null +++ b/doc/examples/cpufreq_governor_efficiency/test.sh @@ -0,0 +1,82 @@ +#!/bin/sh

+set -e

+set_calibration() {

•   calibration.sh
  

+}

+test_efficiency() {

•   FILENAME="results_$RANDOM$$.txt"
  

•   if [ -e /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_governors ]; then
  

•           for i in $(cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_governors); do
  

•                   export gov_$i=$(echo $i)
  

•           done
  

•   else
  

•           echo "cpufreq is not available!"
  

•           exit
  

•   fi
  

•   if [ ! $gov_performance ] ; then
  

•           echo "Can't find performance governor!"
  

•           exit
  

•   fi
  

•   if [ ! $gov_powersave ] ; then
  

•           echo "Can't find powersave governor!"
  

•           exit
  

•   fi
  

•   # Get powersave data
  

•   dvfs.sh powersave $1 $2 $3 > $FILENAME
  

•   powersave=$(cat $FILENAME |sed -n '1p')
  

•   # Get performance data
  

•   dvfs.sh performance $1 $2 $3 > $FILENAME
  

•   performance=$(cat $FILENAME |sed -n '1p')
  

•   if [ $performance -ge $powersave ] ; then
  

•           echo "Error! Probably not input all the cpus in the same frequency domain"
  

•           exit
  

•   fi
  

•   denominator=$(expr $powersave - $performance)
  

•   echo "powersave efficiency: 0%"
  

•   echo "performance efficiency: 100%"
  

•   # Calcuate other governors data
  

•   for gov_next in $gov_conservative $gov_ondemand $gov_cfs; do
  

Why have you restricted the test to these 3 governors ? what about userspace gov ? or interactive governor when available ?

•           if [ "$gov_next" != "" ] ; then
  

•                   dvfs.sh $gov_next $1 $2 $3 > $FILENAME
  

•                   data=$(cat $FILENAME |sed -n '1p');
  

•                   numerator=$(expr $powersave - $data)
  

•                   numerator=$(expr $numerator \* 100)
  

•                   if [ $numerator -lt 0 ] ; then
  

•                           let numerator=0
  

•                   fi
  

•                   data=$(expr $numerator / $denominator)
  

•                   echo "$gov_next efficiency: $data%"
  

•           fi
  

•   done
  

•   rm -f $FILENAME
  

+}

+if [ $# -lt 3 ]; then

•   echo "Usage: ./test.sh <cpus> <runtime> <sleeptime>"
  

•   echo "cpus: number of cpus in the CPU0's frequency domain"
  

•   echo "runtime: running time in ms per loop of the workload pattern"
  

•   echo "sleeptime: sleeping time in ms per loop of the workload pattern"
  

•   echo -e "\nExample: \n\"./test.sh 4 100 1000\" means\nCPU0~CPU3 sharing frequency, \"100ms run + 1000ms sleep\" workload pattern.\n"
  

•   exit
  

+fi

+echo "Frequency domain CPU0~CPU$(expr $1 - 1), run $2ms, sleep $3ms:"

+sleep 1 +PATH=$PATH:. +set_calibration +test_efficiency $1 $(expr $2 * 1000) $(expr $3 * 1000)

You have created several temporary file while executing your script, you should clean them before exiting.

Some governors like ondemand or interactive ones have parameters that modified their responsiveness. How can i set them before testing their efficiency ?

As an example, ondemand governor efficiency moves from nearly 0% to nearly 100% on my chromebook2 (only the quad A15 have been enable for the test) if you change the sampling_rate and the sampling_down_factor when you test short run: 100ms run 1000ms sleep. Default configuration of a governor are not always the best one for the platform

Regards, Vincent

-- 1.9.1