[igt-dev] [PATCH i-g-t 3/4] i915/gem_shrink: Refactor allocation sizing based on available memory

[Chris Wilson]

Quoting Chris Wilson (2020-12-14 16:21:53)
> Quoting Tvrtko Ursulin (2020-12-14 15:57:59)
> > 
> > On 14/12/2020 10:51, Chris Wilson wrote:
> > > Refactor the allocation such that we utilise just enough memory pressure
> > > to invoke the shrinker, and just enough processes to spread across the
> > > CPUs and contend on the shrinker.
> > > 
> > > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > > ---
> > >   tests/i915/gem_shrink.c | 11 ++++++-----
> > >   1 file changed, 6 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/tests/i915/gem_shrink.c b/tests/i915/gem_shrink.c
> > > index 023db8c56..e8a814fe6 100644
> > > --- a/tests/i915/gem_shrink.c
> > > +++ b/tests/i915/gem_shrink.c
> > > @@ -426,6 +426,7 @@ igt_main
> > >       int num_processes = 0;
> > >   
> > >       igt_fixture {
> > > +             const int ncpus = sysconf(_SC_NPROCESSORS_ONLN);
> > >               uint64_t mem_size = intel_get_total_ram_mb();
> > >               int fd;
> > >   
> > > @@ -434,16 +435,16 @@ igt_main
> > >   
> > >               /*
> > >                * Spawn enough processes to use all memory, but each only
> > > -              * uses half the available mappable aperture ~128MiB.
> > > +              * uses half of the available per-cpu memory.
> > >                * Individually the processes would be ok, but en masse
> > >                * we expect the shrinker to start purging objects,
> > >                * and possibly fail.
> > >                */
> > > -             alloc_size = gem_mappable_aperture_size(fd) / 2;
> > > -             num_processes = 1 + (mem_size / (alloc_size >> 20));
> > > +             alloc_size = (mem_size + ncpus - 1) / ncpus / 2;
> > 
> > Div round up with thousands divided by small integers okay, safe on very 
> > old smp boxes. :)
> > 
> > > +             num_processes = ncpus + (mem_size / alloc_size);
> > 
> > Hm, now what does this add up to..
> > 
> > ncpus + mem_size / (mem_size / ncpus / 2) = ... ?
> 
> (ncpus + mem_size / (mem_size / ncpus / 2)) * (mem_size / ncpus / 2)
> 
> mem_size / 2 + mem_size

If we make alloc_size smaller then, say /8, so we get more processes and
less overallocation.
-Chris