lkml.org 
[lkml]   [2021]   [Feb]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] mm, slub: better heuristic for number of cpus when calculating slab order
On Mon, Feb 08, 2021 at 02:41:08PM +0100, Vlastimil Babka wrote:
> When creating a new kmem cache, SLUB determines how large the slab pages will
> based on number of inputs, including the number of CPUs in the system. Larger
> slab pages mean that more objects can be allocated/free from per-cpu slabs
> before accessing shared structures, but also potentially more memory can be
> wasted due to low slab usage and fragmentation.
> The rough idea of using number of CPUs is that larger systems will be more
> likely to benefit from reduced contention, and also should have enough memory
> to spare.
>
> <SNIP>
>
> So this patch tries to determine the best available value without specific arch
> knowledge.
> - num_present_cpus() if the number is larger than 1, as that means the arch is
> likely setting it properly
> - nr_cpu_ids otherwise
>
> This should fix the reported regressions while also keeping the effect of
> 045ab8c9487b for PowerPC systems. It's possible there are configurations where
> num_present_cpus() is 1 during boot while nr_cpu_ids is at the same time
> bloated, so these (if they exist) would keep the large orders based on
> nr_cpu_ids as was before 045ab8c9487b.
>

Tested-by: Mel Gorman <mgorman@techsingularity.net>

Only x86-64 tested, three machines, all showing similar results as would
be expected. One example;

hackbench-process-sockets
5.11.0-rc7 5.11.0-rc7 5.11.0-rc7
vanilla revert-v1r1 vbabka-fix-v1r1
Amean 1 0.3873 ( 0.00%) 0.4060 ( -4.82%) 0.3747 ( 3.27%)
Amean 4 1.3767 ( 0.00%) 0.7700 * 44.07%* 0.7790 * 43.41%*
Amean 7 2.4710 ( 0.00%) 1.2753 * 48.39%* 1.2680 * 48.68%*
Amean 12 3.7103 ( 0.00%) 1.9570 * 47.26%* 1.9470 * 47.52%*
Amean 21 5.9790 ( 0.00%) 2.9760 * 50.23%* 2.9830 * 50.11%*
Amean 30 8.0467 ( 0.00%) 4.0590 * 49.56%* 4.0410 * 49.78%*
Amean 48 12.8180 ( 0.00%) 6.5167 * 49.16%* 6.4070 * 50.02%*
Amean 79 20.5150 ( 0.00%) 10.3580 * 49.51%* 10.3740 * 49.43%*
Amean 110 25.5320 ( 0.00%) 14.0453 * 44.99%* 14.0577 * 44.94%*
Amean 141 32.4170 ( 0.00%) 17.3267 * 46.55%* 17.4977 * 46.02%*
Amean 172 40.0883 ( 0.00%) 21.0360 * 47.53%* 21.1480 * 47.25%*
Amean 203 47.2923 ( 0.00%) 25.2367 * 46.64%* 25.4923 * 46.10%*
Amean 234 55.2623 ( 0.00%) 29.0720 * 47.39%* 29.3273 * 46.93%*
Amean 265 61.4513 ( 0.00%) 33.0260 * 46.26%* 33.0617 * 46.20%*
Amean 296 73.2960 ( 0.00%) 36.6920 * 49.94%* 37.2520 * 49.18%*

Comparing just a revert and the patch

5.11.0-rc7 5.11.0-rc7
revert-v1r1 vbabka-fix-v1r1
Amean 1 0.4060 ( 0.00%) 0.3747 ( 7.72%)
Amean 4 0.7700 ( 0.00%) 0.7790 ( -1.17%)
Amean 7 1.2753 ( 0.00%) 1.2680 ( 0.58%)
Amean 12 1.9570 ( 0.00%) 1.9470 ( 0.51%)
Amean 21 2.9760 ( 0.00%) 2.9830 ( -0.24%)
Amean 30 4.0590 ( 0.00%) 4.0410 ( 0.44%)
Amean 48 6.5167 ( 0.00%) 6.4070 ( 1.68%)
Amean 79 10.3580 ( 0.00%) 10.3740 ( -0.15%)
Amean 110 14.0453 ( 0.00%) 14.0577 ( -0.09%)
Amean 141 17.3267 ( 0.00%) 17.4977 * -0.99%*
Amean 172 21.0360 ( 0.00%) 21.1480 ( -0.53%)
Amean 203 25.2367 ( 0.00%) 25.4923 ( -1.01%)
Amean 234 29.0720 ( 0.00%) 29.3273 ( -0.88%)
Amean 265 33.0260 ( 0.00%) 33.0617 ( -0.11%)
Amean 296 36.6920 ( 0.00%) 37.2520 ( -1.53%)

That's a negligible difference and all but one group (141) was within the
noise. Even for 141, it's very marginal and with the degree of overload
at that group count, it can be ignored.

Thanks!

--
Mel Gorman
SUSE Labs

\
 
 \ /
  Last update: 2021-02-10 15:09    [W:0.075 / U:2.092 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site