From: Holger Kiehl Subject: Re: Performance of ext4 Date: Fri, 20 Jun 2008 08:09:59 +0000 (GMT) Message-ID: References: <18513.345.553912.449710@frecb006361.adech.frec.bull.fr> <20080612131928.GB18229@mit.edu> <20080612180605.GD22481@skywalker> <20080616175408.GF3279@atrey.karlin.mff.cuni.cz> <20080616181353.GA20686@skywalker> <20080619155645.GA8582@mit.edu> Mime-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed Cc: "Aneesh Kumar K.V" , Jan Kara , Solofo.Ramangalahy@bull.net, Nick Dokos , linux-ext4@vger.kernel.org, linux-kernel To: Theodore Tso Return-path: Received: from dwdmx4.dwd.de ([141.38.3.230]:60547 "EHLO dwdmx4.dwd.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751482AbYFTIKB (ORCPT ); Fri, 20 Jun 2008 04:10:01 -0400 Received: from localhost (localhost [127.0.0.1]) by node1.dwd.de (Postfix) with ESMTP id 522CE1F86CF for ; Fri, 20 Jun 2008 08:10:00 +0000 (UTC) Received: from localhost ([127.0.0.1]) by localhost (node1.csg-cluster.lan [127.0.0.1]) (amavisd-new, port 2525) with SMTP id 15118-59 for ; Fri, 20 Jun 2008 08:10:00 +0000 (UTC) In-Reply-To: <20080619155645.GA8582@mit.edu> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Thu, 19 Jun 2008, Theodore Tso wrote: > On Tue, Jun 17, 2008 at 11:42:36AM +0000, Holger Kiehl wrote: >> Note how the size of file results.24033.helena.dwd.de changes from >> 9230 before the test to 8208 bytes after the test. Also note the >> date both have the same timestamp "2008-06-17 04:35". I have made a >> copy of results.24033.helena.dwd.de before the test and compared it >> with that after the test. The file is just truncated by 1022 bytes >> and there is no garbage. > > So the corruption is always a truncation, correct? > Correct. > Did you notice the problem with ext4 w/o the patch queue? > No, without patch queue it I did not not notice the problem. > I have a > suspicion that the problem may have been introduced by the delayed > allocation code, but I don't have hard evidence. When you rerun your > benchmark (which seems to be the closest thing we have to a > reproduction case), it would be interesting to know if the problem > goes away with -o nodelalloc (again, it would localize where we need > to look). > Ok I will do thats soon as I have a system available. Holger