On Fri, Jan 26, 2024 at 8:41 AM T.J. Mercier <tjmercier@xxxxxxxxxx> wrote: > > On Fri, Jan 26, 2024 at 8:34 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote: > > > > On Wed, Jan 24, 2024 at 09:46:23AM -0800, T.J. Mercier wrote: > > > In the meantime, instead of a revert how about changing the batch size > > > geometrically instead of the SWAP_CLUSTER_MAX constant: > > > > > > reclaimed = try_to_free_mem_cgroup_pages(memcg, > > > - min(nr_to_reclaim - > > > nr_reclaimed, SWAP_CLUSTER_MAX), > > > + (nr_to_reclaim - nr_reclaimed)/2, > > > GFP_KERNEL, reclaim_options); > > > > > > I think that should address the overreclaim concern (it was mentioned > > > that the upper bound of overreclaim was 2 * request), and this should > > > also increase the reclaim rate for root reclaim with MGLRU closer to > > > what it was before. > > > > Hahaha. Would /4 work for you? > > > > I genuinely think the idea is worth a shot. /4 would give us a bit > > more margin for error, since the bailout/fairness cutoffs have changed > > back and forth over time. And it should still give you a reasonable > > convergence on MGLRU. > > > > try_to_free_reclaim_pages() already does max(nr_to_reclaim, > > SWAP_CLUSTER_MAX) which will avoid the painful final approach loops > > the integer division would produce on its own. > > > > Please add a comment mentioning the compromise between the two reclaim > > implementations though. > > I'll try it out and get back to you. :) Right, so (nr_to_reclaim - nr_reclaimed)/4 looks pretty good to me: root - full reclaim pages/sec time (sec) pre-0388536ac291 : 68047 10.46 post-0388536ac291 : 13742 inf (reclaim-reclaimed)/4 : 67352 10.51 /uid_0 - 1G reclaim pages/sec time (sec) overreclaim (MiB) pre-0388536ac291 : 258822 1.12 107.8 post-0388536ac291 : 105174 2.49 3.5 (reclaim-reclaimed)/4 : 233396 1.12 -7.4 /uid_0 - full reclaim pages/sec time (sec) pre-0388536ac291 : 72334 7.09 post-0388536ac291 : 38105 14.45 (reclaim-reclaimed)/4 : 72914 6.96 So I'll put up a new patch.
