On Sun, Dec 12, 2021 at 4:37 AM Daniel Latypov <dlatypov@xxxxxxxxxx> wrote:
>
> On Sat, Dec 11, 2021 at 12:49 AM David Gow <davidgow@xxxxxxxxxx> wrote:
> >
> > The --jobs parameter for kunit_tool currently defaults to 8 CPUs,
> > regardless of the number available. For systems with significantly more
> > (or less), this is not as efficient. Instead, default --jobs to the
> > number of CPUs present in the system: while there are as many
> > superstitions as to exactly what the ideal jobs:CPU ratio is, this seems
> > sufficiently sensible to me.
> >
> > Signed-off-by: David Gow <davidgow@xxxxxxxxxx>
>
> Reminder: the unit tests depend on this hard-coded value.
> $ ag '\b8\b' tools/testing/kunit/kunit_tool_test.py
> 422:
> self.linux_source_mock.build_kernel.assert_called_once_with(False, 8,
> '.kunit', None)
> 529:
> self.linux_source_mock.build_kernel.assert_called_once_with(False, 8,
> build_dir, None)
>
Gah: that's what I get for rushing this through at the end of the day.
v2 will have this no-longer hardcoded, but call a get_default_jobs()
function which calculates the default.
> > ---
> > tools/testing/kunit/kunit.py | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/tools/testing/kunit/kunit.py b/tools/testing/kunit/kunit.py
> > index 68e6f461c758..2cb6c7db5683 100755
> > --- a/tools/testing/kunit/kunit.py
> > +++ b/tools/testing/kunit/kunit.py
> > @@ -310,7 +310,7 @@ def add_build_opts(parser) -> None:
> > parser.add_argument('--jobs',
> > help='As in the make command, "Specifies the number of '
> > 'jobs (commands) to run simultaneously."',
> > - type=int, default=8, metavar='jobs')
> > + type=int, default=os.cpu_count(), metavar='jobs')
>
> Just looking for edge cases:
> https://docs.python.org/3/library/os.html#os.cpu_count says
> > Returns None if undetermined
> and
> > This number is not equivalent to the number of CPUs the current process can use. The number of usable CPUs can be obtained with len(os.sched_getaffinity(0))
>
> I assume the None caveat is mainly for other operating systems and
> doubt it'll impact any users.
> The second point is a bit more interesting, but still niche.
> Up to you if you want to use that instead.
>
> Super unscientific comparison (n=1) running all on CPU #0
>
> $ taskset 0x1 ./tools/testing/kunit/kunit.py run --jobs=1
> Elapsed time: ... 155.978s building ...
>
> --jobs=2 (some people swear by the 2x ratio)
> Elapsed time: ... 158.891s building ...
>
> --jobs=8 (Old behavior)
> ...
> Elapsed time: ... 171.448s building
>
> --jobs=32
> Elapsed time: ... 170.765s building ...
>
> So the overhead of j being "too high" isn't that bad and it doesn't
> seem to matter much either way.
>
>
I'll go with len(os.sched_getaffinity(0)) in v2: it is technically
closer to what the intent is here, even if it's not a problem in
general. Putting it in a helper function will make the
len(os.sched_getaffinity(0)) call less unwieldy, IMHO, too.
> >
> > def add_exec_opts(parser) -> None:
> > parser.add_argument('--timeout',
> > --
> > 2.34.1.173.g76aa8bc2d0-goog
> >
Cheers,
-- David
