build performance: bb-matrix on 4-core (BB_NUMBER_THREADS and PARALLEL_MAKE optimization)
I ran the attached bb-matrix.sh on the following system:
CPU (1): Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz
Cores: 4
Threads: 8
Memory: 8186560 kB
OS Disk: INTEL SSDSA2M040G2GC (SSD)
Build Disk: Hitachi HDT721050SLA360 (Spinning Media)
The script runs builds with all combinations of BB_NUMBER_THREADS and PARALLEL_MAKE from 4 through 16.
Once BB_NUMBER_THREADS hit 10, the kernel OOM Killer started killing off tasks and build time tripled. Those runs have been removed the dataset.
All of the runs with PARALLEL_MAKE=10 also failed, for a variety of reasons. See bb-pm-errors.txt for details. For whatever reason, 10 seems to be a bad number. Additional failures were seen at 09-11 and 10-14. These have all been removed from the dat file.
From the remaining results, a clear downward trend in build time is evident with increasing BB_NUMBER_THREADS through 8, while build time mostly increases again with 9 (and dramatically so with 10, not shown). Optimal build time is achieved with BB_NUMBER_THREADS=8.
Along the BB_NUMBER_THREADS=8 line, there is no clear trend with increasing values of PARALLEL_MAKE. Local downward trends appear from 4-7 and from 11-14. Optimal build time occurs with PARALLEL_MAKE=14, however, it only bests PARALLEL_MAKE=7 by 68 seconds.
While optimal build time is achieved with BB=8 and PM=14, a more resource friendly setting of BB=8 and PM=6 yields nearly as good results.
To reproduce the plots and get an interactive view that you can rotate:
$ gnuplot --persist < bb-pm-matrix.plt
Signature · 3 lines
Darren Hart Intel Open Source Technology Center Yocto Project - Linux Kernel
set xlabel "BB_NUMBER_THREADS"
set ylabel "PARALLEL_MAKE"
set zlabel "Build Time (seconds)"
set dgrid3d 13,6
set pm3d at b
set ticslevel 0.8
set term png
set output "plot.png"
splot "bb-pm-runtime-fear-jul6.dat" u 1:2:3 with lines
set view 90,0
set output "plot-bb.png"
replot
set view 90,90
set output "plot-pm.png"
replot
set view 60,30
set term wxt
replot
Quoted 25 lines · Darren Hart
I ran the attached
bb-matrix.shon the following system:CPU (1): Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz
Cores: 4
Threads: 8
Memory: 8186560 kB
OS Disk: INTEL SSDSA2M040G2GC (SSD)
Build Disk: Hitachi HDT721050SLA360 (Spinning Media)The script runs builds with all combinations of
BB_NUMBER_THREADSandPARALLEL_MAKEfrom 4 through 16.Once
BB_NUMBER_THREADShit 10, the kernel OOM Killer started killing off tasks and build time tripled. Those runs have been removed the dataset.All of the runs with PARALLEL_MAKE=10 also failed, for a variety of reasons. See bb-pm-errors.txt for details. For whatever reason, 10 seems to be a bad number. Additional failures were seen at 09-11 and 10-14. These have all been removed from the dat file.
From the remaining results, a clear downward trend in build time is evident with increasing
BB_NUMBER_THREADSthrough 8, while build time mostly increases again with 9 (and dramatically so with 10, not shown). Optimal build time is achieved with BB_NUMBER_THREADS=8.
Quoted 7 lines
Along the BB_NUMBER_THREADS=8 line, there is no clear trend with increasing values of
PARALLEL_MAKE. Local downward trends appear from 4-7 and from 11-14. Optimal build time occurs with PARALLEL_MAKE=14, however, it only bests PARALLEL_MAKE=7 by 68 seconds.While optimal build time is achieved with BB=8 and PM=14, a more resource friendly setting of BB=8 and PM=6 yields nearly as good results.
Thanks Darren, I think those are interesting results.
Is the general advice we should give out therefore to set BB_NUMBER_THREADS = PARALLEL_MAKE = number threads?
I'd love to understand why there is the peak and second dip on the PARALLEL_MAKE curve...
It would also be good to put the script in scripts/contrib.
Cheers,
Richard
Quoted 31 lines · Richard Purdie
Quoted 30 lines · Darren Hart
I ran the attached
bb-matrix.shon the following system:CPU (1): Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz
Cores: 4
Threads: 8
Memory: 8186560 kB
OS Disk: INTEL SSDSA2M040G2GC (SSD)
Build Disk: Hitachi HDT721050SLA360 (Spinning Media)The script runs builds with all combinations of
BB_NUMBER_THREADSandPARALLEL_MAKEfrom 4 through 16.Once
BB_NUMBER_THREADShit 10, the kernel OOM Killer started killing off tasks and build time tripled. Those runs have been removed the dataset.All of the runs with PARALLEL_MAKE=10 also failed, for a variety of reasons. See bb-pm-errors.txt for details. For whatever reason, 10 seems to be a bad number. Additional failures were seen at 09-11 and 10-14. These have all been removed from the dat file.
From the remaining results, a clear downward trend in build time is evident with increasing
BB_NUMBER_THREADSthrough 8, while build time mostly increases again with 9 (and dramatically so with 10, not shown). Optimal build time is achieved with BB_NUMBER_THREADS=8.Along the BB_NUMBER_THREADS=8 line, there is no clear trend with increasing values of
PARALLEL_MAKE. Local downward trends appear from 4-7 and from 11-14. Optimal build time occurs with PARALLEL_MAKE=14, however, it only bests PARALLEL_MAKE=7 by 68 seconds.
Quoted 7 lines
Quoted 2 lines
While optimal build time is achieved with BB=8 and PM=14, a more resource friendly setting of BB=8 and PM=6 yields nearly as good results.
Thanks Darren, I think those are interesting results.
Is the general advice we should give out therefore to set
BB_NUMBER_THREADS=PARALLEL_MAKE= number threads?
I don't think we have enough data to make a general recommendation, but for 4 core systems, I think BB=8 and PM=6 is a good choice. With some additional runs on other hardware, hopefully we can come up with a more general number like BB=2*NR_CORES PM=1.5*NR_CORES (cores not threads).
Quoted 3 lines
I'd love to understand why there is the peak and second dip on the
PARALLEL_MAKEcurve...
Me too! The PM axis plots are very strange and not at all expected. I'm seeing similarly unexpected results on the 12 core machine currently running through a 12-48 bb-matrix run.
Quoted 2 lines
It would also be good to put the script in
scripts/contrib.
Yes, I'll try to get around to that soon... this week. They need a little cleanup I think first.
Signature · 3 lines
Darren Hart Intel Open Source Technology Center Yocto Project - Linux Kernel
Darren/Richard,
Maybe we could instead of hacking hard coded default values (or nothing) into the config file default to something like this:
somehow get the number of CPUs:
CPUS=$(grep ^processor /proc/cpuinfo | wc -l)
echo CPUS=${CPUS}
or
CPUS=getconf _NPROCESSORS_ONLN
echo CPUS_UBUNTU=${CPUS_UBUNTU}
(don't know if the second one will also work with other distros than Ubuntu)
Do some calculation which magic number for BB_NUMBER_THREADS and PARALLEL_MAKE to use:
e.g. what was suggested: BB=2*NR_CORES PM=1.5*NR_CORES
Regards,
Robert
...For every complex problem there is a solution which is simple, neat and wrong. -- H.L. Mencken
My public pgp key is available at:
http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x90320BF1
Quoted 15 lines · Robert Berger
Darren/Richard,
Maybe we could instead of hacking hard coded default values (or nothing) into the config file default to something like this:
somehow get the number of CPUs:
CPUS=$(grep ^processor
/proc/cpuinfo| wc -l)
echo CPUS=${CPUS}or
CPUS=
getconf _NPROCESSORS_ONLN
echo CPUS_UBUNTU=${CPUS_UBUNTU}
Quoted 6 lines
(don't know if the second one will also work with other distros than Ubuntu)
Do some calculation which magic number for
BB_NUMBER_THREADSandPARALLEL_MAKEto use:e.g. what was suggested: BB=2*NR_CORES PM=1.5*NR_CORES
My concern with this is that on larger machines I'm seeing very different optimal multipliers. On my 12 core with a RAID 0 build array, the ideal setting seems to be BB=12 PM=12.
Until we can better characterize the ideal settings, I think we are better off documenting what works for specific systems. Now perhaps we need to do something that caps the number, but that is sure to be wrong in short order as well.
As your signature suggests, the solution to this isn't likely to be simple ;-)
--
Darren
Quoted 15 lines
Regards,
Robert
...For every complex problem there is a solution which is simple, neat and wrong. -- H.L. MenckenMy public pgp key is available at:
http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x90320BF1_______________________________________________
poky mailing list
[email protected]
https://lists.yoctoproject.org/listinfo/poky
Signature · 3 lines
Darren Hart Intel Open Source Technology Center Yocto Project - Linux Kernel
Quoted 27 lines · Darren Hart
Quoted 15 lines · Robert Berger
Darren/Richard,
Maybe we could instead of hacking hard coded default values (or nothing) into the config file default to something like this:
somehow get the number of CPUs:
CPUS=$(grep ^processor
/proc/cpuinfo| wc -l)
echo CPUS=${CPUS}or
CPUS=
getconf _NPROCESSORS_ONLN
echo CPUS_UBUNTU=${CPUS_UBUNTU}Quoted 6 lines
(don't know if the second one will also work with other distros than Ubuntu)
Do some calculation which magic number for
BB_NUMBER_THREADSandPARALLEL_MAKEto use:e.g. what was suggested: BB=2*NR_CORES PM=1.5*NR_CORES
My concern with this is that on larger machines I'm seeing very different optimal multipliers. On my 12 core with a RAID 0 build array, the ideal setting seems to be BB=12 PM=12.
Quoted 7 lines
Until we can better characterize the ideal settings, I think we are better off documenting what works for specific systems. Now perhaps we need to do something that caps the number, but that is sure to be wrong in short order as well.
As your signature suggests, the solution to this isn't likely to be simple ;-)
This may be rather specific to my personal setup, but I use
https://gist.github.com/776390 -- you'll note that you can adjust the scaling factors via variables.
Signature · 5 lines
Christopher Larson clarson at kergoth dot com Founder - BitBake, OpenEmbedded, OpenZaurus Maintainer - Tslib Senior Software Engineer, Mentor Graphics
Quoted 33 lines · Chris Larson
Quoted 32 lines · Darren Hart
Quoted 15 lines · Robert Berger
Darren/Richard,
Maybe we could instead of hacking hard coded default values (or nothing) into the config file default to something like this:
somehow get the number of CPUs:
CPUS=$(grep ^processor
/proc/cpuinfo| wc -l)
echo CPUS=${CPUS}or
CPUS=
getconf _NPROCESSORS_ONLN
echo CPUS_UBUNTU=${CPUS_UBUNTU}Quoted 6 lines
(don't know if the second one will also work with other distros than Ubuntu)
Do some calculation which magic number for
BB_NUMBER_THREADSandPARALLEL_MAKEto use:e.g. what was suggested: BB=2*NR_CORES PM=1.5*NR_CORES
My concern with this is that on larger machines I'm seeing very different optimal multipliers. On my 12 core with a RAID 0 build array, the ideal setting seems to be BB=12 PM=12.
Until we can better characterize the ideal settings, I think we are better off documenting what works for specific systems. Now perhaps we need to do something that caps the number, but that is sure to be wrong in short order as well.
Quoted 6 lines
Quoted 2 lines
As your signature suggests, the solution to this isn't likely to be simple ;-)
This may be rather specific to my personal setup, but I use
https://gist.github.com/776390 -- you'll note that you can adjust the scaling factors via variables.
Something like this would probably be a good improvement - but it will need some sort of step function (of cpu count) for the multipliers. I'm concerned this step function will be tedious to maintain. I suspect the ideal number is also dependent on on build path storage (spinning disk, RAID, SSD, tmpfs, etc.), faster storage can likely benefit from higher thread counts, whereas slower storage just gets more and more bogged down under higher thread counts. I'll have some numbers from the 12 core on Tuesday if I'm extrapolating accurately.
Signature · 3 lines
Darren Hart Intel Open Source Technology Center Yocto Project - Linux Kernel