On Tue 27-10-20 11:35:35, Jann Horn wrote: > On Tue, Oct 27, 2020 at 8:05 AM Michael Kerrisk (man-pages) > <mtk.manpages@xxxxxxxxx> wrote: > > On 10/12/20 4:52 PM, Jann Horn wrote: > > > On Mon, Oct 12, 2020 at 1:49 PM Jann Horn <jannh@xxxxxxxxxx> wrote: > > >> Since 34e55232e59f7b19050267a05ff1226e5cd122a5 (introduced back in > > >> v2.6.34), Linux uses per-thread RSS counters to reduce cache contention on > > >> the per-mm counters. With a 4K page size, that means that you can end up > > >> with the counters off by up to 252KiB per thread. > > > > > > Actually, as Mark Mossberg pointed out to me off-thread, the counters > > > can actually be off by many times more... > > > > So, does your patch to proc.5 need tweaking, or can I just apply as is? > > The "(up to 63 pages per thread)" would have to go, the rest should be correct. > > But as Michal said, if someone volunteers to get rid of this > optimization, maybe we don't need the documentation after all? But > that would probably require actually doing some careful > heavily-multithreaded benchmarking on a big machine with a few dozen > cores, or something like that, so that we know whether this > optimization actually is unimportant enough that we can just get rid > of it... Well, the original micro optimization didn't really come with some solid numbers based on real workloads. Artificial workloads are likely not very representative for this case because any potential counters overhead normally gets dispersed. I think this is the case where the benefit is so unclear that I would simply revert the whole thing and try to tune up for a real life workloads that actually sees a regression. -- Michal Hocko SUSE Labs
