From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753309AbaAQSt2 (ORCPT ); Fri, 17 Jan 2014 13:49:28 -0500 Received: from mx1.redhat.com ([209.132.183.28]:35124 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752733AbaAQStL (ORCPT ); Fri, 17 Jan 2014 13:49:11 -0500 Message-ID: <52D97ACD.10209@redhat.com> Date: Fri, 17 Jan 2014 13:47:41 -0500 From: Rik van Riel User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:24.0) Gecko/20100101 Thunderbird/24.2.0 MIME-Version: 1.0 To: Peter Zijlstra CC: linux-kernel@vger.kernel.org, linux-mm@vger.kernel.org, mingo@redhat.com, chegu_vinod@hp.com, mgorman@suse.de, Rik van Riel Subject: Re: [PATCH 6/6] numa,sched: normalize faults_from stats and weigh by CPU use References: <1389939456-9541-1-git-send-email-riel@redhat.com> <1389939456-9541-7-git-send-email-riel@redhat.com> <20140117095459.GF11314@laptop.programming.kicks-ass.net> <52D94380.5040901@redhat.com> <20140117150409.GJ11314@laptop.programming.kicks-ass.net> In-Reply-To: <20140117150409.GJ11314@laptop.programming.kicks-ass.net> Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 01/17/2014 10:04 AM, Peter Zijlstra wrote: > On Fri, Jan 17, 2014 at 09:51:44AM -0500, Rik van Riel wrote: >> On 01/17/2014 04:54 AM, Peter Zijlstra wrote: >>> On Fri, Jan 17, 2014 at 01:17:36AM -0500, riel@redhat.com wrote: >>>> + /* >>>> + * Normalize the faults_from, so all tasks in a group >>>> + * count according to CPU use, instead of by the raw >>>> + * number of faults. This prevents the situation where >>>> + * the garbage collector totally dominates the stats, >>>> + * and the access patterns of the worker threads are >>>> + * ignored. >>>> + */ >>> >>> Instead of focusing on the one example (GC) here, I would suggest >>> saying something along the lines of: Tasks with little runtime have >>> little over-all impact on throughput and thus their faults are less >>> important. >> >> Thanks for the review, I have updated the comment >> accordingly. >> >> Does anybody else have comments on this series, or should >> I start begging for Acked-by: and Reviewed-by: lines? :) > > I still need to get me head around part of the previous patches, and I > suppose Mel might have a look, give it a few days ;-) Well, you didn't spot the divide by zero that Vinod somehow managed to trigger :) I am also going to add another patch, that prevents the temporary values from "p->numa_faults[i] >>= 1;" becoming visible, by using local variables for those calculations. Expect a v2 soonish. -- All rights reversed