From mboxrd@z Thu Jan 1 00:00:00 1970 From: Dave Chinner Subject: Re: [numa shrinker] 9b17c62382: -36.6% regression on sparse file copy Date: Wed, 15 Jan 2014 11:18:27 +1100 Message-ID: <20140115001827.GO3469@dastard> References: <20140106082048.GA567@localhost> <20140106131042.GA5145@destitution> <20140109025715.GA11984@localhost> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Glauber Costa , Linux Memory Management List , linux-fsdevel@vger.kernel.org, LKML , lkp@linux.intel.com To: Fengguang Wu Return-path: Received: from ipmail05.adl6.internode.on.net ([150.101.137.143]:20281 "EHLO ipmail05.adl6.internode.on.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751068AbaAOASz (ORCPT ); Tue, 14 Jan 2014 19:18:55 -0500 Content-Disposition: inline In-Reply-To: <20140109025715.GA11984@localhost> Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Thu, Jan 09, 2014 at 10:57:15AM +0800, Fengguang Wu wrote: > Hi Dave, > > As you suggested, I added tests for ext4 and btrfs, the results are > the same. > > Then I tried running perf record for 10 seconds starting from 200s. > (The test runs for 410s). I see several warning messages and hope > they do not impact the accuracy too much: > > [ 252.608069] perf samples too long (2532 > 2500), lowering kernel.perf_event_max_sample_rate to 50000 > [ 252.608863] perf samples too long (2507 > 2500), lowering kernel.perf_event_max_sample_rate to 25000 > [ 252.609422] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 1.389 msecs > > Anyway the noticeable perf change are: > > 1d3d4437eae1bb2 9b17c62382dd2e7507984b989 > --------------- ------------------------- > 12.15 ~10% +209.8% 37.63 ~ 2% brickland2/debug2/vm-scalability/300s-btrfs-lru-file-readtwice > 12.88 ~16% +189.4% 37.27 ~ 0% brickland2/debug2/vm-scalability/300s-ext4-lru-file-readtwice > 15.24 ~ 9% +146.0% 37.50 ~ 1% brickland2/debug2/vm-scalability/300s-xfs-lru-file-readtwice > 40.27 +179.1% 112.40 TOTAL perf-profile.cpu-cycles._raw_spin_lock.grab_super_passive.super_cache_count.shrink_slab.do_try_to_free_pages > > 1d3d4437eae1bb2 9b17c62382dd2e7507984b989 > --------------- ------------------------- > 11.91 ~12% +218.2% 37.89 ~ 2% brickland2/debug2/vm-scalability/300s-btrfs-lru-file-readtwice > 12.47 ~16% +200.3% 37.44 ~ 0% brickland2/debug2/vm-scalability/300s-ext4-lru-file-readtwice > 15.36 ~11% +145.4% 37.68 ~ 1% brickland2/debug2/vm-scalability/300s-xfs-lru-file-readtwice > 39.73 +184.5% 113.01 TOTAL perf-profile.cpu-cycles._raw_spin_lock.put_super.drop_super.super_cache_count.shrink_slab > > perf report for 9b17c62382dd2e7507984b989: > > # Overhead Command Shared Object Symbol > # ........ ............... .................. .............................................. > # > 77.74% dd [kernel.kallsyms] [k] _raw_spin_lock > | > --- _raw_spin_lock > | > |--47.65%-- grab_super_passive Oh, it's superblock lock contention, probably caused by an increase in shrinker calls (i.e. per-node rather than global). I think we've seen this before - can you try the two patches from Tim Chen here: https://lkml.org/lkml/2013/9/6/353 https://lkml.org/lkml/2013/9/6/356 If they fix the problem, I'll get them into 3.14 and pushed back to the relevant stable kernels. Cheers, Dave. -- Dave Chinner david@fromorbit.com