TipsAndTricks/InvestigatingBuildTime: Difference between revisions

From Yocto Project
Jump to navigationJump to search
No edit summary
No edit summary
 
(5 intermediate revisions by the same user not shown)
Line 7: Line 7:




The <tt>buildstats</tt> data produce by the class with the same name if included in <tt>USER_CLASSES</tt> variable
The <tt>buildstats</tt> data produced by the class with the same name provides '''process statistics''' for every recipe task.
provides '''process statistics''' for every task. This data  can offer insight of the time spent by the builder (<tt>bitbake</tt>) when
This data  can offer insight of the time spent by the builder (<tt>bitbake</tt>) when
constructing a particular target at process level. This article focus more on profiling than optimization, the former being
constructing a particular target at process level. This article focus more on profiling than optimization, the former being
the base for the latter.
the base for the latter.
Line 37: Line 37:
to either plot the top N recipes-task stat values
to either plot the top N recipes-task stat values


  $ ../scripts/contrib/bb-perf/buildstats-plot.sh -s <BUILDSTATS> -N $N | gnuplot -p
  $ ../scripts/contrib/bb-perf/buildstats-plot.sh -s <BUILDSTATS> -n $N | gnuplot -p


or sum all stats for a particular task for all (world) recipes
or sum stats values per task for all (world) recipes


  $ ../scripts/contrib/bb-perf/buildstats-plot.sh -s <BUILDSTATS> -S | gnuplot -p
  $ ../scripts/contrib/bb-perf/buildstats-plot.sh -s <BUILDSTATS> -S | gnuplot -p


The former produces a more granular plot and second a more general idea where the type is spent per task set.
The former produces a more granular plot and second a more general idea where the type is spent per task set.
=== Setup ===
GPLIMAGE='set terminal png fontscale 1 size 1280,720'
function bsplots() {
  local STAT=$1
  OUT="set output '$STAT.task-recipe.png'"
  ../scripts/contrib/bb-perf/buildstats-plot.sh -s "$STAT" -n 20 > input.gpl
  gnuplot -e "$GPLIMAGE;$OUT" input.gpl
  OUT="set output '$STAT.task.png'"
  ../scripts/contrib/bb-perf/buildstats-plot.sh -s "$STAT" -S > input.gpl
  gnuplot -e "$GPLIMAGE;$OUT" input.gpl
}


=== Results ===
=== Results ===


==== Time Results ====
Preliminar results can be found [[MortyBuildstats|here]]
 
  bsplots "stime:utime"
 
[[File:Stime-utime.task-recipe.png]]
[[File:Stime-utime.task.png]]
 
==== IO Results ====
 
  bsplots "IO rchar:IO wchar"
 
[[File:IOrchar-IOwchar.task-recipe.png]]
[[File:IOrchar-IOwchar.task.png]]

Latest revision as of 21:12, 28 February 2017

Draft: Investigating Build Time

(Placeholder. Demonstrate use of buildstats and pybootchartgui)

Current status: The intention is to use the scripts under scripts/contrib/bb-perf to parse buildstat data and produce useful plots.


The buildstats data produced by the class with the same name provides process statistics for every recipe task. This data can offer insight of the time spent by the builder (bitbake) when constructing a particular target at process level. This article focus more on profiling than optimization, the former being the base for the latter.

Setup

 $ cd poky
 $ git checkout -b morty origin/morty
 $ . oe-init-build-env
 $ bitbake world

Buildstats

buildstats.bbclass stores mainly two set of process' stats: time and IO stats. For the moment, we are currently focusing on the following time data:

  • /proc/[pid]/stat: utime, stime
  • /proc/[pid]/stat: cutime, cstime
  • python resource module: rusage ru_utime, rusage ru_stime
  • python resource module: Child rusage ru_utime, Child rusage ru_stime

and the IO stats:

  • /proc/[pid]/io: IO rchar, IO wchar
  • /proc/[pid]/io: IO read_bytes, IO write_bytes

With this data, the following script (once the bitbake world command has finished) is used to either plot the top N recipes-task stat values

$ ../scripts/contrib/bb-perf/buildstats-plot.sh -s <BUILDSTATS> -n $N | gnuplot -p

or sum stats values per task for all (world) recipes

$ ../scripts/contrib/bb-perf/buildstats-plot.sh -s <BUILDSTATS> -S | gnuplot -p

The former produces a more granular plot and second a more general idea where the type is spent per task set.

Results

Preliminar results can be found here