Content-Type: multipart/mixed; boundary="===============1460907868925031192==" MIME-Version: 1.0 From: Mel Gorman To: lkp@lists.01.org Subject: Re: [mm] f7b5d647946: -3.0% dbench.throughput-MB/sec Date: Tue, 19 Aug 2014 20:11:51 +0100 Message-ID: <20140819191139.GG10146@suse.de> In-Reply-To: <20140819154351.GA2697@localhost> List-Id: --===============1460907868925031192== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Tue, Aug 19, 2014 at 11:43:51PM +0800, Fengguang Wu wrote: > On Tue, Aug 19, 2014 at 03:34:28PM +0100, Mel Gorman wrote: > > On Tue, Aug 19, 2014 at 12:41:34PM +0800, Fengguang Wu wrote: > > > Hi Mel, > > > = > > > We noticed a minor dbench throughput regression on commit > > > f7b5d647946aae1647bf5cd26c16b3a793c1ac49 ("mm: page_alloc: abort fair > > > zone allocation policy when remotes nodes are encountered"). > > > = > > > testcase: ivb44/dbench/100% > > > = > > > bb0b6dffa2ccfbd f7b5d647946aae1647bf5cd26 > > > --------------- ------------------------- > > > 25692 =C2=B1 0% -3.0% 24913 =C2=B1 0% dbench.throughp= ut-MB/sec > > > 6974259 =C2=B1 6% -12.1% 6127616 =C2=B1 0% meminfo.DirectM= ap2M > > > 18.43 =C2=B1 0% -4.6% 17.59 =C2=B1 0% turbostat.RAM_W > > > 9302 =C2=B1 0% -3.6% 8965 =C2=B1 1% time.user_time > > > 1425791 =C2=B1 1% -2.0% 1396598 =C2=B1 0% time.involuntar= y_context_switches > > > = > > > Disclaimer: > > > Results have been estimated based on internal Intel analysis and are = provided > > > for informational purposes only. Any difference in system hardware or= software > > > design or configuration may affect actual performance. > > > = > > = > > DirectMap2M changing is a major surprise and doesn't make sense for this > > machine. > = > The ivb44's hardware configuration is > = > model: Ivytown Ivy Bridge-EP > nr_cpu: 48 > memory: 64G > = > And note that this is an in-memory dbench run, which is why > dbench.throughput-MB/sec is so high. > = Ok, it's a NUMA machine. I expect in this case that prior to the patch more local memory would have been used on node 0 due to the fair zone allocation policy skipping remote nodes. The patch corrects the behaviour of zonelist but the downside is more remote accesses for processes running on node 0. T= he behaviour is correct although not necessarily desirable from a performance point of view. Users should boot with numa_zonelist_order=3Dnode if it's a problem. -- = Mel Gorman SUSE Labs --===============1460907868925031192==--