Content-Type: multipart/mixed; boundary="===============6296626842536075255==" MIME-Version: 1.0 From: Davidlohr Bueso To: lkp@lists.01.org Subject: Re: [mm] c8c06efa8b5: -7.6% unixbench.score Date: Tue, 27 Jan 2015 12:43:45 -0800 Message-ID: <1422391425.4604.7.camel@stgolabs.net> In-Reply-To: <1422344736.24875.47.camel@linux.intel.com> List-Id: --===============6296626842536075255== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Tue, 2015-01-27 at 15:45 +0800, Huang Ying wrote: > Hi, Davidlohr, > = > Any update on this? The latest news I gotten is that you finally > reproduced this. The overall differences and cause of the issue is commit 37e9562453b (locking/rwsem: Allow conservative optimistic spinning when readers have lock). And since the performance issue occurs with writers-only, I'm fine with the degradation, we need to learn to live it. Real use cases will use mutexes instead. Also, the mentioned commit fixes a much more important problem than pathologically calling exec(2). Now, this doesn't mean we cannot do _anything_ about it. I sent a patcshet just a few days ago (sorry, I forgot to Cc you) that tries to minimalize (I've recovered ~75% of the regression) the writer penalties: https://lkml.org/lkml/2015/1/26/34 So all in all we don't really care for writer-only workloads, but hey, while we're at it. Thanks, Davidlohr --===============6296626842536075255==--