From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932078AbXDKSuW (ORCPT ); Wed, 11 Apr 2007 14:50:22 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S932096AbXDKSuW (ORCPT ); Wed, 11 Apr 2007 14:50:22 -0400 Received: from smtp.osdl.org ([65.172.181.24]:56427 "EHLO smtp.osdl.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932078AbXDKSuU (ORCPT ); Wed, 11 Apr 2007 14:50:20 -0400 Date: Wed, 11 Apr 2007 11:49:27 -0700 From: Andrew Morton To: Pavel Emelianov Cc: Paul Menage , Balbir Singh , Linux Kernel Mailing List , devel@openvz.org, Kirill Korotaev , Srivatsa Vaddagiri Subject: Re: [PATCH] Cpu statistics accounting based on Paul Menage patches Message-Id: <20070411114927.1277d7c9.akpm@linux-foundation.org> In-Reply-To: <461CF883.2030308@sw.ru> References: <461CF883.2030308@sw.ru> X-Mailer: Sylpheed version 2.2.7 (GTK+ 2.8.6; i686-pc-linux-gnu) Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Wed, 11 Apr 2007 19:02:27 +0400 Pavel Emelianov wrote: > Provides a per-container statistics concerning the numbers of tasks > in various states, system and user times, etc. Patch is inspired > by Paul's example of the used CPU time accounting. Although this > patch is independent from Paul's example to make it possible playing > with them separately. Why is this actually needed? If userspace has a list of the tasks which are in a particular container, it can run around and add up the stats for those tasks without kernel changes? It's a bit irksome that we have so much accounting of this form in core kernel, yet we have to go and add a completely new implementation to create something which is similar to what we already have. But I don't immediately see a fix for that. Apart from paragraph #1 ;) Should there be linkage between per-container stats and delivery-via-taskstats? I can't think of one, really. You have cpu stats. Later, presumably, we'll need IO stats, MM stats, context-switch stats, number-of-syscall stats, etc, etc. Are we going to reimplement all of those things as well? See paragraph #1! Bottom line: I think we seriously need to find some way of consolidating per-container stats with our present per-task stats. Perhaps we should instead be looking at ways in which we can speed up paragraph #1.