[dpdk-dev] [PATCH] mk: optimize directory dependencies

[Richardson, Bruce]

> -----Original Message-----
> From: Jerin Jacob [mailto:jerin.jacob at caviumnetworks.com]
> Sent: Tuesday, January 24, 2017 12:57 PM
> To: Richardson, Bruce <bruce.richardson at intel.com>
> Cc: Olivier Matz <olivier.matz at 6wind.com>; dev at dpdk.org; Yigit, Ferruh
> <ferruh.yigit at intel.com>; thomas.monjalon at 6wind.com
> Subject: Re: [dpdk-dev] [PATCH] mk: optimize directory dependencies
> 
> On Tue, Jan 24, 2017 at 12:15:06PM +0000, Bruce Richardson wrote:
> > On Tue, Jan 24, 2017 at 05:10:15PM +0530, Jerin Jacob wrote:
> > > On Mon, Jan 23, 2017 at 06:19:13PM +0100, Olivier Matz wrote:
> > > > Before this patch, the management of dependencies between
> > > > directories had several issues:
> > > >
> > > > - the generation of .depdirs, done at configuration is slow: it can
> take
> > > >   more than one minute on some slow targets (usually ~10s on a
> standard
> > > >   PC).
> > > >
> > > > - for instance, it is possible to expressed a dependency like:
> > > >   - app/foo depends on lib/librte_foo
> > > >   - and lib/librte_foo depends on app/bar
> > > >   But this won't work because the directories are traversed with a
> > > >   depth-first algorithm, so we have to choose between doing 'app'
> before
> > > >   or after 'lib'.
> > > >
> > > > - the script depdirs-rule.sh is too complex.
> > > >
> > > > - we cannot use "make -d" for debug, because the output of make is
> used for
> > > >   the generation of .depdirs.
> > > >
> > > > This patch moves the DEPDIRS-* variables in the upper Makefile,
> > > > making the dependencies much easier to calculate. A DEPDIRS
> > > > variable is still used to process library dependencies in LDLIBS.
> > > >
> > > > After this commit, "make config" is almost immediate.
> > > >
> > > > Signed-off-by: Olivier Matz <olivier.matz at 6wind.com>
> > >
> > > Tested both approach on ThunderX. This patch looks better
> > >
> > > Tested-by: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> > >
> > > ➜ [master]GB-2S [dpdk-master] $ time make config
> > > T=arm64-thunderx-linuxapp-gcc Configuration done
> > >
> > > real    0m18.112s
> > > user    0m2.810s
> > > sys     0m0.660s
> > >
> > > ➜ [master]GB-2S [dpdk-master] $ pwclient git-am 19859 Applying patch
> > > #19859 using 'git am'
> > > Description: [dpdk-dev] mk: parallelize make config
> > > Applying: mk: parallelize make config
> > >
> > > ➜ [master]GB-2S [dpdk-master] $ rm -rf build ➜ [master]GB-2S
> > > [dpdk-master] $ time make config T=arm64-thunderx-linuxapp-gcc -j 8
> > > Configuration done
> > >
> > > real    0m2.812s
> > > user    0m3.020s
> > > sys     0m0.870s
> > >
> > > ➜ [master]GB-2S [dpdk-master] $ rm -rf build ➜ [master]GB-2S
> > > [dpdk-master] $ time make config T=arm64-thunderx-linuxapp-gcc -j 16
> > > Configuration done
> > >
> > > real    0m1.748s
> > > user    0m3.040s
> > > sys     0m1.020s
> > >
> > > ➜ [master]GB-2S [dpdk-master] $ rm -rf build ➜ [master]GB-2S
> > > [dpdk-master] $ time make config T=arm64-thunderx-linuxapp-gcc -j 32
> > > Configuration done
> > >
> > > real    0m1.422s
> > > user    0m3.380s
> > > sys     0m1.080s
> > >
> > > ➜ [master]GB-2S [dpdk-master] $ pwclient git-am 19918 Applying patch
> > > #19918 using 'git am'
> > > Description: [dpdk-dev] mk: optimize directory dependencies
> > > Applying: mk: optimize directory dependencies
> > >
> > > ➜ [master]GB-2S [dpdk-master] $ rm -rf build ➜ [master]GB-2S
> > > [dpdk-master] $ time make config T=arm64-thunderx-linuxapp-gcc
> > > Configuration done
> > >
> > > real    0m0.064s
> > > user    0m0.000s
> > > sys     0m0.000s
> > > ➜ [master]GB-2S [dpdk-master] $ rm -rf build ➜ [master]GB-2S
> > > [dpdk-master] $ time make config T=arm64-thunderx-linuxapp-gcc -j 8
> > > Configuration done
> > >
> > > real    0m0.055s
> > > user    0m0.000s
> > > sys     0m0.000s
> > >
> >
> > I agree that Olivier's patch is faster. However, I think I prefer
> > having the library dependencies in the makefiles for the libs
> > themselves rather than up a level. Given that we are only looking at
> > ~2second of a difference here in your tests - assuming -j flag - what
> > is the actual build time differences? My suspicion is that after
> > Ferruh's simpler
> 
> Without patch - 18sec
> With patch -j1 - 18 sec
> With patch -j2 - 9.2 sec
> With patch -j4 - 4.9 sec
> With patch -j8 - 2.8 sec
> With patch -j16 - 1.7 sec
> With patch -j32 - 1.4 sec
> 
> > patch is applied, the config time becomes such a small part of the
> > build, that the extra benefits from Oliviers work is not worth the
> > extra complexity.
> 
> The low-end embedded SoCs (SoC with 2 to 4 cores) will be get benefited
> out of bring this extra complexity.My take is, if it is manageable
> complexity then take the most optimized one.

Ok. Point taken for the lower-core count parts.

Thomas - can at least one of these patches be merged into 17.02, since it will definitely help us developers? [If Olivier's is too big a change at this point in the cycle, can Ferruh's be taken as an intermediate fix for this release before merging Olivier's in 17.05?]

Regards,
/Bruce