From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S262898AbTIQW54 (ORCPT ); Wed, 17 Sep 2003 18:57:56 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S262901AbTIQW54 (ORCPT ); Wed, 17 Sep 2003 18:57:56 -0400 Received: from dyn-ctb-203-221-73-126.webone.com.au ([203.221.73.126]:8452 "EHLO chimp.local.net") by vger.kernel.org with ESMTP id S262898AbTIQW5y (ORCPT ); Wed, 17 Sep 2003 18:57:54 -0400 Message-ID: <3F68E6D2.50204@cyberone.com.au> Date: Thu, 18 Sep 2003 08:57:22 +1000 From: Nick Piggin User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.4) Gecko/20030827 Debian/1.4-3 X-Accept-Language: en MIME-Version: 1.0 To: Mike Sullivan CC: linux-kernel@vger.kernel.org Subject: Re: I/O degredation with AS on 2.6.0-test3 References: In-Reply-To: Content-Type: text/plain; charset=us-ascii; format=flowed Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Mike Sullivan wrote: > >>>Dave Hansen wrote: >>> >>>You might want to try Martin Bligh's diffprofile utility. It's a bit >>>hard to compare 2 500-line profiles without it. >>> >>>ftp://ftp.kernel.org/pub/linux/kernel/people/mbligh/tools/ >>> >>>Also, there have evidently been a few I/O scheduler fixes since -test3. >>>Please retry with -test5. >>> >>> > >>Nick Piggen wrote: >>More important, are they regressions vs. previous kernels with AS? >> > >I've checked 2.5.69mm9, 2.5.75, 2.6.0test3, and 2.6.0test5 and they all >show a >significant degradation in IO performance (>40%) when using the as >scheduler >compared to the deadline scheduler. > >Configuration: IBM x440 8way with hypertheading enabled, 16GB RAM, >4 QLA2310 FC adapters attached to 2 FastT900 controllers >(112 disks total). The 112 physical disks are striped as 8 raid 0 >logical disks. There are a total of 40 raw devices setup (5 per raid0 >disk). > >I've reattached excerpts from the database test for vmstat and >the readprofile diffs created from Martin's diffprofile utility >(diffprofile readp.as readp.dl). > Thanks Mike, I'll get a patch for you to try soon. I'm a bit busy ATM.