On Tue 03-10-23 01:03:53, Yosry Ahmed wrote: > On Tue, Oct 3, 2023 at 12:57 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > On Mon 25-09-23 10:11:05, Yosry Ahmed wrote: > > > On Mon, Sep 25, 2023 at 6:50 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > > > > > On Fri 22-09-23 17:57:38, Yosry Ahmed wrote: > > > > > While working on adjacent code [1], I realized that the values passed > > > > > into memcg_rstat_updated() to keep track of the magnitude of pending > > > > > updates is consistent. It is mostly in pages, but sometimes it can be in > > > > > bytes or KBs. Fix that. > > > > > > > > What kind of practical difference does this change make? Is it worth > > > > additional code? > > > > > > As explained in patch 2's commit message, the value passed into > > > memcg_rstat_updated() is used for the "flush only if not worth it" > > > heuristic. As we have discussed in different threads in the past few > > > weeks, unnecessary flushes can cause increased global lock contention > > > and/or latency. > > > > > > Byte-sized paths (percpu, slab, zswap, ..) feed bytes into the > > > heuristic, but those are interpreted as pages, which means we will > > > flush earlier than we should. This was noticed by code inspection. How > > > much does this matter in practice? I would say it depends on the > > > workload: how many percpu/slab allocations are being made vs. how many > > > flushes are requested. > > > > > > On a system with 100 cpus, 25M of stat updates are needed for a flush > > > usually, but ~6K of slab/percpu updates will also (mistakenly) cause a > > > flush. > > > > This surely depends on workload and that is understandable. But it would > > be really nice to provide some numbers for typical workloads which > > exercise slab heavily. > > If you have a workload in mind I can run it and see how many flushes > we get with/without this patch. The first thing that pops into my head > is creating a bunch of empty files but I don't know if that's the best > thing to get numbers from. Let me remind you that you are proposing a performance optimization and such a change requires some numbers to actually show it is benefitial. There are cases where the resulting code is clearly an improvement and the performance benefit is just a nice side effect. I do not consider this to be the case. The whole thing is quite convoluted even without a better precision you are proposing. And let me be clear, I am not opposing your patch but I would rather see it based on more than just hand waving. -- Michal Hocko SUSE Labs
