From: Vivek Goyal Subject: Re: [patch,rfc v2] ext3/4: enhance fsync performance when using cfq Date: Wed, 7 Apr 2010 17:46:31 -0400 Message-ID: <20100407214631.GL3206@redhat.com> References: Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: "Theodore Ts'o" , jens.axboe@oracle.com, linux-ext4@vger.kernel.org, linux-kernel@vger.kernel.org To: Jeff Moyer Return-path: Received: from mx1.redhat.com ([209.132.183.28]:14919 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752640Ab0DGVql (ORCPT ); Wed, 7 Apr 2010 17:46:41 -0400 Content-Disposition: inline In-Reply-To: Sender: linux-ext4-owner@vger.kernel.org List-ID: On Wed, Apr 07, 2010 at 05:18:12PM -0400, Jeff Moyer wrote: > Hi again, > > So, here's another stab at fixing this. This patch is very much an RFC, > so do not pull it into anything bound for Linus. ;-) For those new to > this topic, here is the original posting: http://lkml.org/lkml/2010/4/1/344 > > The basic problem is that, when running iozone on smallish files (up to > 8MB in size) and including fsync in the timings, deadline outperforms > CFQ by a factor of about 5 for 64KB files, and by about 10% for 8MB > files. From examining the blktrace data, it appears that iozone will > issue an fsync() call, and will have to wait until it's CFQ timeslice > has expired before the journal thread can run to actually commit data to > disk. > > The approach below puts an explicit call into the filesystem-specific > fsync code to yield the disk so that the jbd[2] process has a chance to > issue I/O. This bring performance of CFQ in line with deadline. > > There is one outstanding issue with the patch that Vivek pointed out. > Basically, this could starve out the sync-noidle workload if there is a > lot of fsync-ing going on. I'll address that in a follow-on patch. For > now, I wanted to get the idea out there for others to comment on. > > Thanks a ton to Vivek for spotting the problem with the initial > approach, and for his continued review. > Thanks Jeff. Conceptually this appraoch makes lot of sense to me. Higher layers explicitly telling CFQ not to idle/yield the slice. My firefox timing test is perfoming much better now. real 0m15.957s user 0m0.608s sys 0m0.165s real 0m12.984s user 0m0.602s sys 0m0.148s real 0m13.057s user 0m0.624s sys 0m0.145s So we got to take care of two issues now. - Make it work with dm/md devices also. Somehow shall have to propogate this yield semantic down the stack. - The issue of making sure we don't yield if we are servicing sync-noidle tree and there is other IO going on which relies on sync-noidle tree idling (as you have already mentioned). [..] > +static void cfq_yield(struct request_queue *q) > +{ > + struct cfq_data *cfqd = q->elevator->elevator_data; > + struct cfq_io_context *cic; > + struct cfq_queue *cfqq; > + unsigned long flags; > + > + cic = cfq_cic_lookup(cfqd, current->io_context); > + if (!cic) > + return; > + > + spin_lock_irqsave(q->queue_lock, flags); > + > + /* > + * This is primarily called to ensure that the long synchronous > + * time slice does not prevent other I/O happenning (like journal > + * commits) while we idle waiting for it. Thus, check to see if the > + * current cfqq is the sync cfqq for this process. > + */ > + cfqq = cic_to_cfqq(cic, 1); > + if (!cfqq) > + goto out_unlock; > + > + if (cfqd->active_queue != cfqq) > + goto out_unlock; > + > + cfq_log_cfqq(cfqd, cfqq, "yielding queue"); > + > + __cfq_slice_expired(cfqd, cfqq, 1); > + __blk_run_queue(cfqd->queue); I think it would be good if we also check that cfqq is empty or not. If cfqq is not empty we don't want to give up slice? But this is a minor point. At least in case of fsync, we seem to be waiting for all the iozone request to finish and then calling yield. So cfqq should be empty at this point of time. Vivek