Content-Type: multipart/mixed; boundary="===============6230071117890711400==" MIME-Version: 1.0 From: Minchan Kim To: lkp@lists.01.org Subject: Re: [mm] 755d6edc1a: will-it-scale.per_process_ops -4.1% regression Date: Tue, 06 Aug 2019 20:00:24 +0900 Message-ID: <20190806110024.GA32615@google.com> In-Reply-To: <20190806080415.GG11812@dhcp22.suse.cz> List-Id: --===============6230071117890711400== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Tue, Aug 06, 2019 at 10:04:15AM +0200, Michal Hocko wrote: > On Tue 06-08-19 15:05:47, kernel test robot wrote: > > Greeting, > > = > > FYI, we noticed a -4.1% regression of will-it-scale.per_process_ops due= to commit: > = > I have to confess I cannot make much sense from numbers because they > seem to be too volatile and the main contributor doesn't stand up for > me. Anyway, regressions on microbenchmarks like this are not all that > surprising when a locking is slightly changed and the critical section > made shorter. I have seen that in the past already. I guess if it's multi process workload. The patch will give more chance to be scheduled out so TLB miss ratio would be bigger than old. I see it's natural trade-off for latency vs. performance so only thing I could think is just increase threshold from 32 to 64 or 128? > = > That being said I would still love to get to bottom of this bug rather > than play with the lock duration by a magic. In other words > http://lkml.kernel.org/r/20190730125751.GS9330(a)dhcp22.suse.cz Yes, if we could remove mark_page_accessed there, it would be best. I added a commen in the thread. --===============6230071117890711400==--