Content-Type: multipart/mixed; boundary="===============1584712054725236715==" MIME-Version: 1.0 From: Jens Axboe To: lkp@lists.01.org Subject: Re: [block] 34b48db66e0: +3291.6% iostat.sde.wrqm/s Date: Wed, 21 Jan 2015 22:37:42 -0700 Message-ID: <54C08CA6.8050101@fb.com> In-Reply-To: <1421889689.6126.45.camel@intel.com> List-Id: --===============1584712054725236715== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On 01/21/2015 06:21 PM, Huang Ying wrote: > FYI, we noticed the below changes on > > commit 34b48db66e08ca1c1bc07cf305d672ac940268dc ("block: remove artifical= max_hw_sectors cap") > > testbox/testcase/testparams: lkp-ws02/fileio/600s-100%-1HDD-btrfs-64G-102= 4f-seqrewr-sync > > c2661b806092d8ea 34b48db66e08ca1c1bc07cf305 > ---------------- -------------------------- > %stddev %change %stddev > \ | \ > 47176 =C2=B1 2% -67.3% 15406 =C2=B1 4% softirqs.BLOCK > 1110 =C2=B1 44% -51.0% 544 =C2=B1 35% sched_debug.cpu#= 8.curr->pid > 22 =C2=B1 33% -48.9% 11 =C2=B1 43% sched_debug.cpu#= 1.cpu_load[0] > 91 =C2=B1 43% +125.0% 204 =C2=B1 32% sched_debug.cfs_= rq[4]:/.blocked_load_avg > 17 =C2=B1 46% -65.2% 6 =C2=B1 31% sched_debug.cfs_= rq[1]:/.runnable_load_avg > 105 =C2=B1 43% +102.6% 213 =C2=B1 32% sched_debug.cfs_= rq[4]:/.tg_load_contrib > 163 =C2=B1 35% +62.6% 265 =C2=B1 27% sched_debug.cfs_= rq[16]:/.blocked_load_avg > 183 =C2=B1 29% +51.4% 277 =C2=B1 26% sched_debug.cfs_= rq[16]:/.tg_load_contrib > 1411 =C2=B1 31% -42.5% 812 =C2=B1 32% sched_debug.cpu#= 6.curr->pid > 57565068 =C2=B1 15% +66.8% 96024066 =C2=B1 17% cpuidle.C1E-NHM.= time > 94625 =C2=B1 9% -32.5% 63893 =C2=B1 4% cpuidle.C3-NHM.u= sage > 200 =C2=B1 14% -22.8% 155 =C2=B1 24% sched_debug.cfs_= rq[8]:/.tg_load_contrib > 244 =C2=B1 33% -39.0% 149 =C2=B1 11% sched_debug.cfs_= rq[6]:/.blocked_load_avg > 265 =C2=B1 31% -38.4% 163 =C2=B1 9% sched_debug.cfs_= rq[6]:/.tg_load_contrib > 4959 =C2=B1 9% -18.2% 4058 =C2=B1 1% slabinfo.kmalloc= -128.active_objs > 4987 =C2=B1 9% -18.6% 4058 =C2=B1 1% slabinfo.kmalloc= -128.num_objs > 19 =C2=B1 8% -19.7% 15 =C2=B1 14% sched_debug.cpu#= 0.cpu_load[1] > 662307 =C2=B1 7% -12.6% 579108 =C2=B1 3% cpuidle.C6-NHM.u= sage > 3028 =C2=B1 7% -12.3% 2656 =C2=B1 1% slabinfo.ext4_ex= tent_status.num_objs > 3028 =C2=B1 7% -12.3% 2656 =C2=B1 1% slabinfo.ext4_ex= tent_status.active_objs > 4.87 =C2=B1 0% +3291.6% 165.07 =C2=B1 0% iostat.sde.wrqm/s > 1006 =C2=B1 0% +120.3% 2216 =C2=B1 0% iostat.sde.avgrq= -sz So these two above tells us that we are doing way more write merges per = second, and that the average request size has roughly doubled from 1006 = to 2216 - both are excellent news. > 466 =C2=B1 0% +115.9% 1007 =C2=B1 0% iostat.sde.await > 466 =C2=B1 0% +115.9% 1007 =C2=B1 0% iostat.sde.w_awa= it Service time roughly doubled, must be mostly stream time. > 301 =C2=B1 0% -52.7% 142 =C2=B1 0% iostat.sde.w/s About half the number of writes completed, but from the stats above, = those writes are more than double. 1006 * 301 < 2216 * 142, so again, = this looks good. > 2230 =C2=B1 2% -8.2% 2048 =C2=B1 2% vmstat.system.in And a nice reduction in irq rate, also nice. Way less software irqs from = the first few lines, also a win. -- = Jens Axboe --===============1584712054725236715==--