From: Josef Bacik Subject: Re: transaction batching performance & multi-threaded synchronous writers Date: Tue, 15 Jul 2008 16:43:56 -0400 Message-ID: <20080715204356.GE30311@unused.rdu.redhat.com> References: <487B7B9B.3020001@gmail.com> <20080715183904.GC30311@unused.rdu.redhat.com> <20080715201010.GD30311@unused.rdu.redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Ric Wheeler , linux-ext4@vger.kernel.org, adilger@sun.com To: Josef Bacik Return-path: Received: from mx1.redhat.com ([66.187.233.31]:42034 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1762977AbYGOVCw (ORCPT ); Tue, 15 Jul 2008 17:02:52 -0400 Content-Disposition: inline In-Reply-To: <20080715201010.GD30311@unused.rdu.redhat.com> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Tue, Jul 15, 2008 at 04:10:10PM -0400, Josef Bacik wrote: > On Tue, Jul 15, 2008 at 02:39:04PM -0400, Josef Bacik wrote: > > On Mon, Jul 14, 2008 at 12:15:23PM -0400, Ric Wheeler wrote: > > > > > > Here is a pointer to the older patch & some results: > > > > > > http://www.spinics.net/lists/linux-fsdevel/msg13121.html > > > > > > I will retry this on some updated kernels, but would not expect to see a > > > difference since the code has not been changed ;-) > > > > > > > Ok here are the numbers with the original idea I had proposed. > > > > type threads base patch speedup > > sata 1 17.9 17.3 0.97 > > sata 2 33.2 34.2 1.03 > > sata 4 58.4 63.6 1.09 > > sata 8 78.8 80.8 1.03 > > sata 16 94.4 97.6 1.16 > > > > ram 1 2394.4 1878.0 0.78 > > ram 2 989.6 2041.1 2.06 > > ram 4 1466.1 3201.8 2.18 > > ram 8 1858.1 3362.8 1.81 > > ram 16 3008.0 3227.7 1.07 > > > > I've got to find a fast disk array to test this with, but the ramdisk results > > make me happy, though they were kind of irratic, so I think the fast disk array > > will be a more stable measure of how well this patch does, but it definitely > > doesn't hurt the slow case, and brings stability to the fast case. Thanks much, > > > > Hmm talking with ric I should just leave the single thread stuff alone. That > removes the slight speed regression seen above. Thanks, > Here are the results with the single thread stuff put back in and with 250HZ instead of 1000HZ from before type threads base patch sata 1 21.8 21.6 sata 2 26.2 34.6 sata 4 48.0 58.0 sata 8 70.4 75.2 sata 16 89.6 101.1 ram 1 2505.4 2422.0 ram 2 463.8 3462.3 ram 4 330.4 3653.9 ram 8 995.1 3592.4 ram 16 1335.2 3806.5 Thanks, Josef