Suvayu Ali writes:
Hi Sam,
On Sun, May 31, 2015 at 08:05:44PM -0400, Sam Varshavchik wrote:
> Suvayu Ali writes:
> >
> >How does rpmbulid decide how many threads to run? I think it looks at
> >the local machine and decides. I also recall some variable called
> >RPMBUILD_NCPU or something like that. That would explain why it only
> >uses as many threads as the local machine can handle.
> >
> >Hope this helps,
>
> Sadly, no it doesn't. As I wrote above:
>
> >> >> >> ps shows the maximum number of "/usr/bin/distcc
/usr/bin/g++
> >[options]"
> >> >> >> processes running locally,
>
> rpmbuild uses the _smp_mflags macro to pass the -j parameter to make, and I
> have it correctly configured to kick off ten parallel processes, to match
> the maximum number of ten processes distcc is configured (4 local, 6
> remote). Yet, despite the fact that a manual build distributes the compiles
> correctly, with rpmbuild distcc throttles the number of concurrent parallel
> jobs that it kicks off to four, and runs them locally.
I think you misunderstood me, I wasn't clear either. I saw you set
_smp_mflags, what I mean is rpmbuild still does some magic underneath.
I.e. after you have set that flag, afaik, rpmbuild will fiddle with the
value and pass an "appropriate" -j<n> flag to Make in the regular case.
It makes this choice based on how many threads it can run locally. At
least that's what I think it does.
Okay I did some digging. This is what rpmbuild does with _smp_mflags
(output from rpmbuild --showrc).
-14: _smp_mflags %([ -z "$RPM_BUILD_NCPUS" ] \
&& RPM_BUILD_NCPUS="`/usr/bin/getconf _NPROCESSORS_ONLN`"; \
ncpus_max=%{?_smp_ncpus_max}; \
if [ -n "$ncpus_max" ] && [ "$ncpus_max" -gt 0 ]
&&
[ "$RPM_BUILD_NCPUS" -gt "$ncpus_max" ]; then
RPM_BUILD_NCPUS="$ncpus_max";
fi; \
if [ "$RPM_BUILD_NCPUS" -gt 1 ]; then echo
"-j$RPM_BUILD_NCPUS"; fi)
-14: _smp_ncpus_max 16
I don't have the time to check the above snippet, maybe you can take a
look? Hope this time we follow each other. If this doesn't help, I'm
out of ideas :-(.
That's nice, except, as I mentioned, I override all of this nonsense in
my .rpmmacros, and replace it with a simple
%_smp_mflags -j 10
Once again: the correct number of distcc processes get started. But they
throttle each other, and only four, at a time, kick off a compile on
localhost, the configured distcc max job setting for localhosts, with the
others waiting until the earlier ones finish, before starting the real g++.
distcc does that on its own. It's configured for a maximum number of
concurrent jobs in /etc/distcc/hosts, and if more than that are started,
they throttle each other.
I believe I already explained this, but I'm having some apparent difficulty
carrying across that, for some reason, distcc ends up ignoring everything
but the localhost setting, in /etc/distcc/hosts, when it's started from
rpmbuild. That's where the problem is. Obviously rpmbuild has something to
do with it, but it's not the direct setting. rpmbuild's %_smp_mflags is
configured correctly, and it's doing it's job of kicking off make -j 10, as
it should be.
I was hoping that someone came across this glitch before, and knew the root
cause. I guess not, and I suppose I'll have to strace the whole thing, and
see if that uncovers any clues.