From: Jens Axboe Subject: Re: [GIT PULL] Ext3 latency fixes Date: Sat, 4 Apr 2009 19:36:42 +0200 Message-ID: <20090404173642.GG5178@kernel.dk> References: <1238742067-30814-1-git-send-email-tytso@mit.edu> <20090404135719.GA9812@mit.edu> <20090404151649.GE5178@kernel.dk> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Theodore Tso , Linux Kernel Developers List , Ext4 Developers List To: Linus Torvalds Return-path: Received: from brick.kernel.dk ([93.163.65.50]:46459 "EHLO kernel.dk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752406AbZDDRgo (ORCPT ); Sat, 4 Apr 2009 13:36:44 -0400 Content-Disposition: inline In-Reply-To: Sender: linux-ext4-owner@vger.kernel.org List-ID: On Sat, Apr 04 2009, Linus Torvalds wrote: > > > On Sat, 4 Apr 2009, Linus Torvalds wrote: > > > > > > > On Sat, 4 Apr 2009, Jens Axboe wrote: > > > > > > Big nack on this patch. Ted, this is EXACTLY where I told you we saw big > > > write regressions (sqlite performance drops by a factor of 4-5). Do a > > > git log on fs/buffer.c and see the original patch (which does what your > > > patch does) and the later revert. No idea why you are now suggestion > > > making that exact change?! > > > > Jens, if I can re-create the 'fsync' times (I haven't yet), then the > > default scheduler _will_ be switched to AS. > > Btw, that patch is "obviously correct". > > That write we're submitting is very much a synchronous write. After all, > the code is literally > > ret = submit_bh(WRITE, bh); > wait_on_buffer(bh); > > and it just doesn't get any more synchronous than that. If we don't start > the IO immediately (since we're _waiting_ for it immediately), we're > broken. > > Now, if we need to fix some mysql throughput issue as a result, then I'd > suggest that we look at whether "sync_dirty_buffer()" is sometimes called > when it doesn't need to be od (b) whether perhaps the unplugging behavior > is simply buggy in some other way. > > But Ted's patch makes so much sense on a purely conceptual level, that > when you look at the patch, you should almost not even need to see the > performance numbers to know it's right. But together with the numbers Ted > posted, it's a total no-brainer. CFQ is clearly broken here, and it's > pretty clear that apparently CFQ has been tuned (improperly) purely for > throughput. I agree, hence I previously wrote and submitted an IDENTICAL patch. I'll go and test things, not sure why you think that AS and CFQ perform very differently here, to my knowledge no such postings exist for this test case. And I'll state again that if they do, of course I'll look into that and fix it. One thing that may be of concern is the immediate unplug, but on the other hand, we do an immediate wait which would unplug anyway. So if we just look at the sqlite regression, I'm sure it'll be pretty easy to pin point with some blktrace data. -- Jens Axboe