On Fri, 20 Apr 2012, Ian Lance Taylor wrote:
Marc Glisse <marc.glisse@xxxxxxxx> writes:
building gcc in parallel (make -j): works ok
installing gcc in parallel: not necessary (used to fail regularly, but
I haven't tried recently)
running the testsuite in parallel: not so parallel
On an 8-core PC, for a c,c++ build, make -k -j check leaves the
computer at least 60% idle (I looked at top periodically), and for a
huge part of the time it uses at most 2 cores (and often only 1).
Since it looks like the testsuite is supposed to run in parallel, I
wonder if I am doing it the wrong way, if there are options to pass or
a different command to run?
The horrible test harness that we use, DejaGNU, does not support running
tests in parallel. Running them in parallel is implemented as a hack in
which the tests are manually split up into sets, and the various sets
are passed to DejaGNU in parallel.
Yes, I saw your blog on the subject.
So what you are seeing is that
manually splitting up tests does not lead to optimal use of a parallel
build.
There is no simple fix, but as the actual splitting was done several
years ago it is possible that new tests have overloaded some of the
subsets. Perhaps simply resplitting the tests would lead to better
parallelism. I don't know. If you want to investigate, look at
check_gcc_parallelize and friends in gcc/Makefile.in.
If the only problem was unbalanced targets, I should see a high load at
the beginning that becomes lower and lower as only a few long-running
targets remain. Here, the load does go down, but even at the beginning it
is low, which I find suspicious.
If I figure something out, I'll post it, but it seems hard...
Thanks for your answer,
--
Marc Glisse
