From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751722Ab1ITM7g (ORCPT ); Tue, 20 Sep 2011 08:59:36 -0400 Received: from mx2.parallels.com ([64.131.90.16]:59935 "EHLO mx2.parallels.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751295Ab1ITM7f (ORCPT ); Tue, 20 Sep 2011 08:59:35 -0400 Message-ID: <4E788E08.2020707@parallels.com> Date: Tue, 20 Sep 2011 09:58:48 -0300 From: Glauber Costa User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:6.0.2) Gecko/20110906 Thunderbird/6.0.2 MIME-Version: 1.0 To: Peter Zijlstra CC: , , , , , , Subject: Re: [PATCH 6/9] Include idle and iowait fields in cpuacct References: <1316030695-19826-1-git-send-email-glommer@parallels.com> <1316030695-19826-7-git-send-email-glommer@parallels.com> <1316510490.11841.15.camel@twins> <4E7888B8.3080809@parallels.com> <1316523502.13664.5.camel@twins> In-Reply-To: <1316523502.13664.5.camel@twins> Content-Type: text/plain; charset="UTF-8"; format=flowed Content-Transfer-Encoding: 7bit X-Originating-IP: [201.82.134.226] Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 09/20/2011 09:58 AM, Peter Zijlstra wrote: > On Tue, 2011-09-20 at 09:36 -0300, Glauber Costa wrote: >> On 09/20/2011 06:21 AM, Peter Zijlstra wrote: >>> On Wed, 2011-09-14 at 17:04 -0300, Glauber Costa wrote: >>>> These are slightly different from the others though: >>>> (note to reviewers: might be better to put those in a separate >>>> array?) >>>> >>>> Since idle/iowait are a property of the system - by definition, >>>> no process from any cgroup is running when the system is idle, >>>> they are system wide. So what these fields really mean, are baselines >>>> for when the cgroup was created. It allows the cgroup to start >>>> counting idle/iowait from 0. >>> >>> Alternatively you can make iowait based on nr_uninterruptible per cgroup >>> and count all ticks _this_ cgroup was idle. >> You think? >> >> Humm,humm... maybe... >> iowait can indeed be seen as a process group characteristic. I was >> mainly concerned about overhead here, specially for the idle case: > > The overhead of accounting per cgroup nr_uninterruptible is the worst I > think, that's in the sleep/wakeup paths. > >> If we are idle, there is no task context we can draw from, since the >> task in the cpu is the idle task. So we end up having to touch all >> cgroups... Or am I missing something? >> >> Sounds expensive. > > Count the total number of ticks on the cpu (I think we already have > that) and subtract the number of ticks in this cgroup (I think we also > already have that), which should yield: number of ticks not in this > cgroup, aka number of ticks this cgroup was idle. No , no... remember steal time.