From mboxrd@z Thu Jan 1 00:00:00 1970 From: Frank Mayhar Subject: Re: Results of my VFS scaling evaluation. Date: Mon, 11 Oct 2010 11:47:28 -0700 Message-ID: <1286822848.29899.305.camel@bobble.smo.corp.google.com> References: <1286580739.3153.57.camel@bobble.smo.corp.google.com> <20101009003842.GH30846@shell> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Cc: linux-fsdevel@vger.kernel.org, linux-mm@kvack.org, mrubin@google.com To: Valerie Aurora Return-path: Received: from smtp-out.google.com ([74.125.121.35]:45638 "EHLO smtp-out.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752044Ab0JKSrh (ORCPT ); Mon, 11 Oct 2010 14:47:37 -0400 Received: from hpaq1.eem.corp.google.com (hpaq1.eem.corp.google.com [172.25.149.1]) by smtp-out.google.com with ESMTP id o9BIlZ5I026519 for ; Mon, 11 Oct 2010 11:47:35 -0700 Received: from vws4 (vws4.prod.google.com [10.241.21.132]) by hpaq1.eem.corp.google.com with ESMTP id o9BIkUKd024213 for ; Mon, 11 Oct 2010 11:47:34 -0700 Received: by vws4 with SMTP id 4so442288vws.26 for ; Mon, 11 Oct 2010 11:47:34 -0700 (PDT) In-Reply-To: <20101009003842.GH30846@shell> Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Fri, 2010-10-08 at 20:38 -0400, Valerie Aurora wrote: > On Fri, Oct 08, 2010 at 04:32:19PM -0700, Frank Mayhar wrote: > > > > Before going into details of the test results, however, I must say that > > the most striking thing about Nick's work how stable it is. In all of > > :D > > > the work I've been doing, all the kernels I've built and run and all the > > tests I've run, I've run into no hangs and only one crash, that in an > > area that we happen to stress very heavily, for which I posted a patch, > > available at > > http://www.kerneltrap.org/mailarchive/linux-fsdevel/2010/9/27/6886943 > > The crash involved the fact that we use cgroups very heavily, and there > > was an oversight in the new d_set_d_op() routine that failed to clear > > flags before it set them. > > I honestly can't stand the d_set_d_op() patch (testing flags instead > of d_op->op) because it obfuscates the code in such a way that leads > directly to this kind of bug. I don't suppose you could test the > performance effect of that specific patch and see how big of a > difference it makes? I do kind of understand why he did it but you're right that it makes things a bit error-prone. Unfortunately I'm not in a position at the moment to do a lot more testing and analysis. I'll try to find some spare time in which to do some more testing of both this and Dave Chinner's tree, but no promises. -- Frank Mayhar Google Inc.