From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail1.windriver.com (mail1.windriver.com [147.11.146.13]) by mail.openembedded.org (Postfix) with ESMTP id 2FF2D770F4 for ; Wed, 7 Sep 2016 12:44:39 +0000 (UTC) Received: from ALA-HCA.corp.ad.wrs.com (ala-hca.corp.ad.wrs.com [147.11.189.40]) by mail1.windriver.com (8.15.2/8.15.1) with ESMTPS id u87CiaSW029873 (version=TLSv1 cipher=AES128-SHA bits=128 verify=FAIL); Wed, 7 Sep 2016 05:44:36 -0700 (PDT) Received: from server.local (147.11.117.229) by ALA-HCA.corp.ad.wrs.com (147.11.189.40) with Microsoft SMTP Server id 14.3.248.2; Wed, 7 Sep 2016 05:44:35 -0700 To: Markus Lehtonen , Richard Purdie , openembedded-core References: <1473240442.20226.111.camel@linuxfoundation.org> <86d45370-2472-5be2-d1c9-b0e44bd53291@windriver.com> <1473251611.10544.9.camel@linux.intel.com> From: Bruce Ashfield Message-ID: <9bbebe09-5e7b-0125-460e-54c5ecb4c95a@windriver.com> Date: Wed, 7 Sep 2016 08:44:34 -0400 User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.11; rv:45.0) Gecko/20100101 Thunderbird/45.2.0 MIME-Version: 1.0 In-Reply-To: <1473251611.10544.9.camel@linux.intel.com> Subject: Re: Speed regression in the 4.8 kernel? X-BeenThere: openembedded-core@lists.openembedded.org X-Mailman-Version: 2.1.12 Precedence: list List-Id: Patches and discussions about the oe-core layer List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Wed, 07 Sep 2016 12:44:41 -0000 Content-Type: text/plain; charset="utf-8"; format=flowed Content-Transfer-Encoding: 7bit On 2016-09-07 8:33 AM, Markus Lehtonen wrote: > On Wed, 2016-09-07 at 07:56 -0400, Bruce Ashfield wrote: >> On 2016-09-07 5:27 AM, Richard Purdie wrote: >>> Hi Bruce, >>> >>> I deliberately spaced out the merges of various things so we could get >>> performance measurements of the system as it happened. Unfortunately >>> the 4.8 kernel appears to regress the kernel build time quite >>> significantly: >>> >>> The raw data: >>> >>> ypperf02,master:9428b19a7dd1d265d9f3211004391abe33ea0224,uninative-1.3 >>> -414 >>> -g9428b19,1:01:32,4:21.16,1:00:32,2:10.86,0:16.19,0:11.21,0:01.20,5:34. >>> 73,26701616,6445160,1477762,5446116 >>> ypperf02,master:9428b19a7dd1d265d9f3211004391abe33ea0224,uninative-1.3 >>> -414 >>> -g9428b19,1:04:14,4:23.82,1:00:40,2:13.70,0:16.18,0:11.28,0:01.22,5:45. >>> 48,26697516,6445724,1478048,5446490 >>> ypperf02,master:b9d90ace005597ba35b59adcd8106a1c52e40c1a,uninative-1.3 >>> -438 >>> -gb9d90ac,1:03:16,7:22.13,1:02:46,2:16.60,0:16.32,0:11.04,0:01.21,5:42. >>> 11,30852876,10550952,1707255,5912282 >>> ypperf02,master:f7ca989ddc6d470429b547647d3fbaad83a982d9,uninative-1.3 >>> -446 >>> -gf7ca989,1:04:42,7:29.05,1:03:04,2:19.71,0:16.21,0:11.05,0:01.24,5:52. >>> 83,30845748,10551316,1707615,5912122 >>> >>> which shows the time for "bitbake virtual/kernel -c cleansstate; time >>> bitbake virtual/kernel" goes from 4:20 to 7:22. The disk footprint of >>> the build went from 26GB to 30GB. The build with rm_work size went from >>> 6.4GB to 10.5GB. The overall build time went up 2-3 minutes which looks >>> like the increased kernel build time. The increased time may well be >>> from the increased data being generated/processed. >> >> Is it the actual compile itself ? What's the trick to actually get >> individual task >> >> For the disk footprint, I can check the refs in the tree and purge >> anything that may be dangling. That being said, I can't do that to >> the repository on the git server, so we may need someone that can >> issue a git gc directly on the repository. >> >>> >>> We can't ship M3 with this much of a performance degradation and >>> increased space usage :(. Any idea what changed? >> >> Nope. I can only focus on one thing at a time. I was worried about >> a functionally correct kernel (which I still am) and haven't looked >> at anything in the peripheral yet. >> >> If I can get individual task timings, I can look at it more. >> >> I'm seeing significantly faster meta data phases, etc, so I'm interested >> to know if this is purely in the build steps. > > In my own test setup I'm seeing similar increase in kernel build time. > > Comparing the buildstats of kernel builds from before and after the 4.8 I > got the following numbers (these are cpu times consumed by the tasks > > TASK ABSDIFF RELDIFF CPUTIME1 CPUTIME2 > do_package +17.5s +133.1% 13.1s -> 30.6s > do_deploy +19.9s +1449.4% 1.4s -> 21.3s > do_package_write_rpm +86.8s +714.7% 12.1s -> 98.9s > do_compile_kernelmodules +139.4s +35.9% 388.2s -> 527.6s > do_compile +201.1s +30.0% 670.3s -> 871.4s ok. So as long as the significant increases aren't in do_kernel_metadata or do_patch (those two I've measured), we are dealing with something in the kernel build itself. I can deal with the footprint by inspecting the repo, but Kbuild is a tougher nut to crack. Out of curiosity, is the 4.4 kernel on master showing a significantly shorter build time ? I'm assuming the before is the 4.4 kernel .. but I just wanted to be sure, so we can rule out something else that might be pathologically reacting to the sheer amount of I/O in a kernel build. Bruce > > > I haven't tried to analyze what is causing this yet, though. > > > Thanks, > Markus >