From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S265242AbTLaVEq (ORCPT ); Wed, 31 Dec 2003 16:04:46 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S265249AbTLaVEq (ORCPT ); Wed, 31 Dec 2003 16:04:46 -0500 Received: from mail3.bluewin.ch ([195.186.1.75]:64656 "EHLO mail3.bluewin.ch") by vger.kernel.org with ESMTP id S265242AbTLaVEo (ORCPT ); Wed, 31 Dec 2003 16:04:44 -0500 Date: Wed, 31 Dec 2003 22:03:54 +0100 From: Roger Luethi To: Jens Axboe Cc: Thomas Molina , William Lee Irwin III , Andy Isaacson , Kernel Mailing List Subject: Re: 2.6.0 performance problems Message-ID: <20031231210354.GA9804@k3.hellgate.ch> Mail-Followup-To: Jens Axboe , Thomas Molina , William Lee Irwin III , Andy Isaacson , Kernel Mailing List References: <20031230012551.GA6226@k3.hellgate.ch> <20031230132145.B32120@hexapodia.org> <20031230194051.GD22443@holomorphy.com> <20031230222403.GA8412@k3.hellgate.ch> <20031231101741.GA4378@k3.hellgate.ch> <20031231112119.GB3089@suse.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20031231112119.GB3089@suse.de> X-Operating-System: Linux 2.6.0-test11 on i686 X-GPG-Fingerprint: 92 F4 DC 20 57 46 7B 95 24 4E 9E E7 5A 54 DC 1B X-GPG: 1024/80E744BD wwwkeys.ch.pgp.net User-Agent: Mutt/1.5.4i Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Wed, 31 Dec 2003 12:21:19 +0100, Jens Axboe wrote: > > Thanks. 2.5.39 alone will do, actually. I'm just curious how far the > > similarity between qsbench and bk export goes. > > 2.5.39 is when the deadline io scheduler was merged. How do you define > the qsbench suckiness? 2.5.39 was the biggest regression for qsbench (fixed later, most notably in 2.5.41). 2.5.39 was a significant improvement for efax ("fixed" in 2.5.43). All I'm doing here is reading the graph I posted at: http://hellgate.ch/bench/thrash.tar.gz For the systematic testing, I used "qsbench -p 4 -m 96" on a 256 MB machine. This allowed the kernel to achieve high performance with unfairness -- that's what 2.4 does: One process dominates all others and finishes very early, taking away most of the memory pressure. The spike for qsbench in 2.5.39 remains if only one process is forked (-p1 -m 384), though. I asked for the bk export numbers with 2.5.39 because I'm curious how close to qsbench the behavior really is. > Do you have numbers for 2.4.x and 2.6.1-rc with > the various io schedulers (it would be interesting to see stock, > elevator=deadline, and elevator=noop). I planned to compare the io schedulers in 2.6.0 anyway. Do you expect different numbers for a recent bk snapshot? Roger