From: Al Boldi Subject: Re: CFQ will be the new default IO scheduler - why? Date: Tue, 25 Jul 2006 07:56:49 +0300 Message-ID: <200607250756.49071.a1426z@gawab.com> References: <200607241857.43399.a1426z@gawab.com> <1153758806.3043.55.camel@laptopd505.fenrus.org> Mime-Version: 1.0 Content-Type: text/plain; charset="windows-1256" Content-Transfer-Encoding: 7bit Cc: linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org Return-path: Received: from [212.70.37.245] ([212.70.37.245]:15368 "EHLO raad.intranet") by vger.kernel.org with ESMTP id S932455AbWGYE43 (ORCPT ); Tue, 25 Jul 2006 00:56:29 -0400 To: Arjan van de Ven In-Reply-To: <1153758806.3043.55.camel@laptopd505.fenrus.org> Content-Disposition: inline Sender: linux-fsdevel-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org Arjan van de Ven wrote: > > > Should there be a default scheduler per filesystem? As some > > > filesystems may perform better/worse with one over another? > > > > It's currently perDevice, and should probably be extended to perMount. > > Hi, Hi! > per mount is going to be "not funny". I assume the situation you are > aiming for is the "3 partitions on a disk, each wants its own elevator". > The way the kernel currently works is that IO requests the filesystem > does are first flattened into an IO for the entire device (eg the > partition mapping is done) and THEN the IO scheduler gets involved to > schedule the IO on a per disk basis. IC. That probably explains why concurrent io-procs have such a hard time getting through to the disk. They probably just hang in the flatting phase, waiting for something to take care of their requests. > The 2.4 kernel did this the other way around, and it was really a bad > idea (no fairness, less optimal scheduling all around due to less > visibility into what the disk is really doing, several hardware > properties such as TCQ depth that affect scheduling IOs are truely per > disk not per partition etc etc) > > So I don't think it's likely that per mount is really an option right > now.. Probably true as it stands right now, but extending io-sched semantics to be filesystem aware in the "flattening/partition mapping" phase could improve performance a lot. Thanks! -- Al