From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id ; Tue, 1 Apr 2003 07:05:33 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id ; Tue, 1 Apr 2003 07:05:33 -0500 Received: from proxy.povodiodry.cz ([62.77.115.11]:11935 "HELO pc11.op.pod.cz") by vger.kernel.org with SMTP id ; Tue, 1 Apr 2003 07:05:31 -0500 From: "Vitezslav Samel" Date: Tue, 1 Apr 2003 14:16:53 +0200 To: Alan Cox , Andre Hedrick Cc: linux-kernel@vger.kernel.org Subject: Re: [IDE SiI680] throughput drop to 1/4 Message-ID: <20030401121653.GA1313@pc11.op.pod.cz> Mail-Followup-To: Alan Cox , Andre Hedrick , linux-kernel@vger.kernel.org References: <20030324072910.GA16596@pc11.op.pod.cz> <20030324072910.GA16596@pc11.op.pod.cz> <200303241213.h2OCD6u21467@devserv.devel.redhat.com> <20030325070605.GA26860@pc11.op.pod.cz> <20030326065650.GA10282@pc11.op.pod.cz> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20030326065650.GA10282@pc11.op.pod.cz> User-Agent: Mutt/1.4i Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Wed, Mar 26, 2003 at 07:56:50AM +0100, Vitezslav Samel wrote: > On Tue, Mar 25, 2003 at 08:06:05AM +0100, Vitezslav Samel wrote: > > On Mon, Mar 24, 2003 at 07:13:06AM -0500, Alan Cox wrote: > > > > Recently I tried to figure out in 2.5.65, why throughput on my disk which > > > > hangs on Silicon Image 680 dropped to 1/4 compared to 2.4.21-pre5, but didn't > > > > found anything useful. Are there any known issues with this driver? > > > > > > The same code in both cases. Its quite likely the problem is higher up in > > > the block or filesystem layer. It might also be a general IDE layer bug > > > > > > What does performance look like on your other disk between > > > 2.4.21pre/2.5.65 ? > > > > Will test it as I come back home (it's on my home machine). But I think this > > performance drop is only on the SiI680 interface. > > Tested today: on the other disk on integrated interface (piix) is throughput > exactly the same on both kernels. I put my eyes on /etc/rc.d/* and found hdparm doing some magic; then I tried to fiddle with its params and came to this: the problem lied in fs readahead (hdparm -a xxx). Tried various numbers: fs readahead throughput in 2.5.66 throughput in 2.4.21-pre6 ----------------------------------------------------------------- 0 11.94 MB/sec 40.00 MB/sec 8 12.01 MB/sec 38.79 MB/sec 16 12.04 MB/sec 40.25 MB/sec 24 15.22 MB/sec 39.51 MB/sec 32 15.17 MB/sec 40.00 MB/sec 40 15.20 MB/sec 39.02 MB/sec 48 15.26 MB/sec 40.00 MB/sec 56 17.60 MB/sec 40.25 MB/sec 64 16.03 MB/sec 40.00 MB/sec 64 15.95 MB/sec 40.25 MB/sec 72 17.32 MB/sec 40.00 MB/sec 80 18.34 MB/sec 40.25 MB/sec 88 18.72 MB/sec 40.25 MB/sec 96 16.04 MB/sec 39.26 MB/sec 104 16.68 MB/sec 40.25 MB/sec 112 18.25 MB/sec 39.02 MB/sec 120 18.42 MB/sec 40.00 MB/sec 128 19.17 MB/sec 39.26 MB/sec 136 26.67 MB/sec 40.00 MB/sec 144 39.31 MB/sec 39.51 MB/sec 152 39.58 MB/sec 39.75 MB/sec 160 39.68 MB/sec 40.25 MB/sec 168 39.98 MB/sec 40.25 MB/sec 176 39.85 MB/sec 40.51 MB/sec 184 40.18 MB/sec 40.00 MB/sec So I disabled setting fs readahead (the default values are O.K.). Seems like setting fs readahead doesn't make any change on 2.4 kernels. Cheers, Vita