Re: [PATCH v5] mm/mempolicy: Weighted Interleave Auto-tuning

[Gregory Price <gourry@gourry.net>]

On Mon, Feb 10, 2025 at 04:39:41PM -0800, Andrew Morton wrote:
> On Mon, 10 Feb 2025 00:36:16 -0500 Gregory Price <gourry@gourry.net> wrote:
> 
> > On Fri, Feb 07, 2025 at 06:20:09PM -0800, Andrew Morton wrote:
> > > On Fri,  7 Feb 2025 12:13:35 -0800 Joshua Hahn <joshua.hahnjy@gmail.com> wrote:
> > > 
> > > Leading to... how do we know that this patch makes the kernel better?
> > 
> > Just focusing on this question:
> > 
> > The default behavior of weighted interleave without this patch is
> > equivalent to normal interleave.  This provides a differentiation
> > out-of-the box, and that's just a better experience.
> > 
> > We may find the default values / calculations need tweaking in the
> > future, but this gives us a good starting point.  Anecdotally, I've
> > seen an "optimal" distribution of 10:1 based on the numbers run
> > sub-optimally compared to 7:1 or 13:1 (but better than default mempol).
> 
> How was this optimality measured/observed?
> 

TL;DR: We used MLC to observe highest sustained bandwidth.

Unfortunately I can't post exact numbers at this time.

To simplify the results - HMAT reported bandwidth often drifted
+/- 10% compared to real observed bandwidth. So the distributions
produced by auto-configuration were mildly off - but not by enough
to cause performance degredation, we still saw higher sustained
bandwidth.

When testing the manual configurations I saw that changing from the
auto-selected values to a few ticks in one direction or the other
resulted in *slightly* better results. Not too surprising.

So as long has hardware doesn't lie horrifically, which might be a
tall ask, auto config has a good shot at giving a decent default.

~Gregory