Jonathan Cottrill wrote:
Bruce Dubbs wrote:
Jonathan Cottrill wrote:
As mentioned in Chapter 6, the Perl tests take quite a bit of time when
run
with "make -k test". The testing framework used by Perl doesn't respect
the -j
option to make, so the tests never run in parallel.
There's another way to run the Perl test suite, using the test_harness
make
target and the TEST_JOBS variable; for example:
TEST_JOBS=8 make test_harness
(I picked 8 in this case based on trial and error and load averages on my
quad-core system.) On my system, the test time went from 10m6s to 2m24s.
The docs (https://perldoc.perl.org/perlhack.html#Parallel-tests) do point
out
a caveat that could make this less than ideal for LFS builders: There are
some
tests that supposedly become flaky when run in parallel, like dist/IO/t/
io_dir.t. However, I've run the suite several times this way, and have yet
to
see any new failures (the expected Compress-Raw-Zlib and IO-Compress
failures
still occur, of course). dist/IO/t/io_dir.t is always reported as "ok".
A slight oddity of running this way is that the output is a bit different.
Skipped tests have additional information about why they're skipped, and
the
final report on test failures is more detailed.
At about 2.5 SBU, I do not think the test time for perl is excessive. I
think we can just leave this as is.
I have added notes to libtool and autoconf on how to speed up tests.
Makes sense to me. Thanks!
I'm specifically working on improving build/test times for packages, now; if I
find other things, is there a threshold in SBU where it's worth mentioning on
this list? Don't want to spam people, but also happy to pass on my findings
when they might be helpful.
Posting here is fine. It does get picked up and can be found with various
search engines.
-- Bruce
--
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page
