Content-Type: multipart/mixed; boundary="===============6392486640228849090==" MIME-Version: 1.0 From: Peter Zijlstra To: lkp@lists.01.org Subject: Re: [mutex] 871a6bb4916: -1.8% will-it-scale.per_process_ops, -98.3% will-it-scale.time.voluntary_context_switches, +209.6% will-it-scale.time.involuntary_context_switches Date: Mon, 16 Feb 2015 16:17:51 +0100 Message-ID: <20150216151751.GE5029@twins.programming.kicks-ass.net> In-Reply-To: <1423986414.5538.8.camel@intel.com> List-Id: --===============6392486640228849090== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Sun, Feb 15, 2015 at 03:46:54PM +0800, Huang Ying wrote: > FYI, we noticed the below changes on > = > git://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git locking/co= re > commit 871a6bb4916fef3123b6ff749b0dc82680fb0d2a ("mutex: In mutex_spin_on= _owner(), return true when owner changes") > = > = > testbox/testcase/testparams: wsm/will-it-scale/performance-writeseek3 > = > e07e0d4cb0c4bfe8 871a6bb4916fef3123b6ff749b = > ---------------- -------------------------- = > %stddev %change %stddev > \ | \ = > 24972759 =C2=B1 2% -98.3% 417134 =C2=B1 9% will-it-scale.tim= e.voluntary_context_switches > 2223 =C2=B1 49% +209.6% 6884 =C2=B1 10% will-it-scale.tim= e.involuntary_context_switches > 542 =C2=B1 32% +91.3% 1037 =C2=B1 0% will-it-scale.tim= e.system_time > 186 =C2=B1 30% +86.3% 347 =C2=B1 0% will-it-scale.tim= e.percent_of_cpu_this_job_got > 26.11 =C2=B1 5% -22.7% 20.18 =C2=B1 2% will-it-scale.tim= e.user_time > 0.09 =C2=B1 1% -18.2% 0.07 =C2=B1 1% will-it-scale.sca= lability > 783528 =C2=B1 0% -1.8% 769550 =C2=B1 0% will-it-scale.per= _process_ops > 12.27 =C2=B1 10% +492.7% 72.73 =C2=B1 1% perf-profile.cpu-= cycles.mutex_optimistic_spin.__mutex_lock_slowpath.mutex_lock.generic_file_= write_iter.new_sync_write > 3.22 =C2=B1 26% +1718.0% 58.50 =C2=B1 1% perf-profile.cpu-= cycles.osq_lock.mutex_optimistic_spin.__mutex_lock_slowpath.mutex_lock.gene= ric_file_write_iter Do you have the results for more/different performance tests for this commit? Jason mentioned seeing +5% on one (fserver). So if, for multiple tests, we see an avg improvement, we might trade that for this one regression. If otoh we see a net negative, we'll have to go do something about this. --===============6392486640228849090==--