[dpdk-dev] [PATCH] validate_abi: build faster by augmenting make with job count

[Neil Horman]

On Thu, Jul 21, 2016 at 02:09:19PM +0000, Wiles, Keith wrote:
> 
> > On Jul 21, 2016, at 8:54 AM, Neil Horman <nhorman at tuxdriver.com> wrote:
> > 
> > On Wed, Jul 20, 2016 at 10:32:28PM +0000, Wiles, Keith wrote:
> >> 
> >>> On Jul 20, 2016, at 3:16 PM, Neil Horman <nhorman at tuxdriver.com> wrote:
> >>> 
> >>> On Wed, Jul 20, 2016 at 07:47:32PM +0000, Wiles, Keith wrote:
> >>>> 
> >>>>> On Jul 20, 2016, at 12:48 PM, Neil Horman <nhorman at redhat.com> wrote:
> >>>>> 
> >>>>> On Wed, Jul 20, 2016 at 07:40:49PM +0200, Thomas Monjalon wrote:
> >>>>>> 2016-07-20 13:09, Neil Horman:
> >>>>>>> From: Neil Horman <nhorman at redhat.com>
> >>>>>>> 
> >>>>>>> John Mcnamara and I were discussing enhacing the validate_abi script to build
> >>>>>>> the dpdk tree faster with multiple jobs.  Theres no reason not to do it, so this
> >>>>>>> implements that requirement.  It uses a MAKE_JOBS variable that can be set by
> >>>>>>> the user to limit the job count.  By default the job count is set to the number
> >>>>>>> of online cpus.
> >>>>>> 
> >>>>>> Please could you use the variable name DPDK_MAKE_JOBS?
> >>>>>> This name is already used in scripts/test-build.sh.
> >>>>>> 
> >>>>> Sure
> >>>>> 
> >>>>>>> +if [ -z "$MAKE_JOBS" ]
> >>>>>>> +then
> >>>>>>> +	# This counts the number of cpus on the system
> >>>>>>> +	MAKE_JOBS=`lscpu -p=cpu | grep -v "#" | wc -l`
> >>>>>>> +fi
> >>>>>> 
> >>>>>> Is lscpu common enough?
> >>>>>> 
> >>>>> I'm not sure how to answer that.  lscpu is part of the util-linux package, which
> >>>>> is part of any base install.  Theres a variant for BSD, but I'm not sure how
> >>>>> common it is there.
> >>>>> Neil
> >>>>> 
> >>>>>> Another acceptable default would be just "-j" without any number.
> >>>>>> It would make the number of jobs unlimited.
> >>>> 
> >>>> I think the best is just use -j as it tries to use the correct number of jobs based on the number of cores, right?
> >>>> 
> >>> -j with no argument (or -j 0), is sort of, maybe what you want.  With either of
> >>> those options, make will just issue jobs as fast as it processes dependencies.
> >>> Dependent on how parallel the build is, that can lead to tons of waiting process
> >>> (i.e. more than your number of online cpus), which can actually hurt your build
> >>> time.
> >> 
> >> I read the manual and looked at the code, which supports your statement. (I think I had some statement on stack overflow and the last time I believe anything on the internet :-) I have not seen a lot of differences in compile times with -j on my system. Mostly I suspect it is the number of paths in the dependency, cores and memory on the system.
> >> 
> >> I have 72 lcores or 2 sockets, 18 cores per socket. Xeon 2.3Ghz cores.
> >> 
> >> $ export RTE_TARGET=x86_64-native-linuxapp-gcc 
> >> 
> >> $ time make install T=${RTE_TARGET}
> >> real	0m59.445s user	0m27.344s sys	0m7.040s
> >> 
> >> $ time make install T=${RTE_TARGET} -j
> >> real	0m26.584s user	0m14.380s sys	0m5.120s
> >> 
> >> # Remove the x86_64-native-linuxapp-gcc
> >> 
> >> $ time make install T=${RTE_TARGET} -j 72
> >> real	0m23.454s user	0m10.832s sys	0m4.664s
> >> 
> >> $ time make install T=${RTE_TARGET} -j 8
> >> real	0m23.812s user	0m10.672s sys	0m4.276s
> >> 
> >> cd x86_64-native-linuxapp-gcc
> >> $ make clean
> >> $ time make
> >> real	0m28.539s user	0m9.820s sys	0m3.620s
> >> 
> >> # Do a make clean between each build.
> >> 
> >> $ time make -j
> >> real	0m7.217s user	0m6.532s sys	0m2.332s
> >> 
> >> $ time make -j 8
> >> real	0m8.256s user	0m6.472s sys	0m2.456s
> >> 
> >> $ time make -j 72
> >> real	0m6.866s user	0m6.184s sys	0m2.216s
> >> 
> >> Just the real time numbers in the following table.
> >> 
> >> processes     real Time   depdirs
> >>     no -j             59.4s        Yes
> >>       -j 8             23.8s        Yes
> >>      -j 72            23.5s        Yes
> >>        -j               26.5s        Yes
> >> 
> >>     no -j             28.5s         No
> >>       -j 8               8.2s         No
> >>      -j 72              6.8s         No
> >>        -j                 7.2s         No
> >> 
> >> Looks like the depdirs build time on my system:
> >> $ make clean -j
> >> $ rm .depdirs
> >> $ time make -j
> >> real	0m23.734s user	0m11.228s sys	0m4.844s
> >> 
> >> About 16 seconds, which is not a lot of savings. Now the difference from no -j to -j is a lot, but the difference between -j and -j <cpu_count> is not a huge saving. This leads me back to over engineering the problem when ‘-j’ would work just as well here.
> >> 
> >> Even on my MacBook Pro i7 system the difference is not that much 1m8s without depdirs build for -j in a VirtualBox with all 4 cores 8G RAM. Compared to 1m13s with -j 4 option.
> >> 
> >> I just wonder if it makes a lot of sense to use cpuinfo in this given case if it turns out to be -j works with the 80% rule?
> >> 
> > It may, but that seems to be reason to me to just set DPDK_MAKE_JOBS=0, and
> > you'll get that behavior
> 
> Just to be sure, ‘make -j 0’ is not a valid argument to the -j option. It looks like you have to do ‘-j’ or ‘-j N’ or no option where N != 0
> 
> I think we just use -j which gets us the 80% rule and the best performance without counting cores.
> 
Thats odd, specifying 0 works for me.  If it doesn't for you, specify $MAX_INT
or some other huge number would be comparable

Neil

> > 
> > Neil
> > 
> >> On some other project with a lot more files like the FreeBSD or Linux distro, yes it would make a fair amount of real time difference.
> >> 
> >> Keith
> >> 
> >>> 
> >>> While its fine in los of cases, its not always fine, and with this
> >>> implementation you can still opt in to that behavior by setting DPDK_MAKE_JOBS=0
> >>> 
> >>> Neil
>