Message-ID: <20110221164909.GG6584@quack.suse.cz>
Date:	Mon, 21 Feb 2011 17:49:09 +0100
From:	Jan Kara <jack@...e.cz>
To:	Shaohua Li <shaohua.li@...el.com>
Cc:	"Shi, Alex" <alex.shi@...el.com>, Jan Kara <jack@...e.cz>,
	Corrado Zoccolo <czoccolo@...il.com>,
	Vivek Goyal <vgoyal@...hat.com>,
	"tytso@....edu" <tytso@....edu>,
	"jaxboe@...ionio.com" <jaxboe@...ionio.com>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"Chen, Tim C" <tim.c.chen@...el.com>
Subject: Re: [performance bug] kernel building regression on 64 LCPUs
 machine

On Tue 15-02-11 09:10:01, Shaohua Li wrote:
> On Mon, 2011-02-14 at 10:25 +0800, Shi, Alex wrote:
> > On Sun, 2011-02-13 at 02:25 +0800, Corrado Zoccolo wrote:
> > > On Sat, Feb 12, 2011 at 10:21 AM, Alex,Shi <alex.shi@...el.com> wrote:
> > > > On Wed, 2011-01-26 at 16:15 +0800, Li, Shaohua wrote:
> > > >> On Thu, Jan 20, 2011 at 11:16:56PM +0800, Vivek Goyal wrote:
> > > >> > On Wed, Jan 19, 2011 at 10:03:26AM +0800, Shaohua Li wrote:
> > > >> > > add Jan and Theodore to the loop.
> > > >> > >
> > > >> > > On Wed, 2011-01-19 at 09:55 +0800, Shi, Alex wrote:
> > > >> > > > Shaohua and I tested kernel building performance on latest kernel. and
> > > >> > > > found it is drop about 15% on our 64 LCPUs NHM-EX machine on ext4 file
> > > >> > > > system. We find this performance dropping is due to commit
> > > >> > > > 749ef9f8423054e326f. If we revert this patch or just change the
> > > >> > > > WRITE_SYNC back to WRITE in jbd2/commit.c file. the performance can be
> > > >> > > > recovered.
> > > >> > > >
> > > >> > > > iostat report show with the commit, read request merge number increased
> > > >> > > > and write request merge dropped. The total request size increased and
> > > >> > > > queue length dropped. So we tested another patch: only change WRITE_SYNC
> > > >> > > > to WRITE_SYNC_PLUG in jbd2/commit.c, but nothing effected.
> > > >> > > since WRITE_SYNC_PLUG doesn't work, this isn't a simple no-write-merge issue.
> > > >> > >
> > > >> >
> > > >> > Yep, it does sound like reduce write merging. But moving journal commits
> > > >> > back to WRITE, then fsync performance will drop as there will be idling
> > > >> > introduced between fsync thread and journalling thread. So that does
> > > >> > not sound like a good idea either.
> > > >> >
> > > >> > Secondly, in presence of mixed workload (some other sync read happening)
> > > >> > WRITES can get less bandwidth and sync workload much more. So by
> > > >> > marking journal commits as WRITES you might increase the delay there
> > > >> > in completion in presence of other sync workload.
> > > >> >
> > > >> > So Jan Kara's approach makes sense that if somebody is waiting on
> > > >> > commit then make it WRITE_SYNC otherwise make it WRITE. Not sure why
> > > >> > did it not work for you. Is it possible to run some traces and do
> > > >> > more debugging that figure out what's happening.
> > > >> Sorry for the long delay.
> > > >>
> > > >> Looks fedora enables ccache by default. While our kbuild test is on ext4 disk
> > > >> but rootfs is on ext3 where ccache cache files live. Jan's patch only covers
> > > >> ext4, maybe this is the reason.
> > > >> I changed jbd to use WRITE for journal_commit_transaction. With the change and
> > > >> Jan's patch, the test seems fine.
> > > > Let me clarify the bug situation again.
> > > > With the following scenarios, the regression is clear.
> > > > 1, ccache_dir setup at rootfs that format is ext3 on /dev/sda1; 2,
> > > > kbuild on /dev/sdb1 with ext4.
> > > > but if we disable the ccache, only do kbuild on sdb1 with ext4. There is
> > > > no regressions whenever with or without Jan's patch.
> > > > So, problem focus on the ccache scenario, (from fedora 11, ccache is
> > > > default setting).
> > > >
> > > > If we compare the vmstat output with or without ccache, there is too
> > > > many write when ccache enabled. According the result, it should to do
> > > > some tunning on ext3 fs.
> > > Is ext3 configured with data ordered or writeback?
> > 
> > The ext3 on sda and ext4 on sdb are both used 'ordered' mounting mode. 
> > 
> > > I think ccache might be performing fsyncs, and this is a bad workload
> > > for ext3, especially in ordered mode.
> > > It might be that my patch introduced a regression in ext3 fsync
> > > performance, but I don't understand how reverting only the change in
> > > jbd2 (that is the ext4 specific journaling daemon) could restore it.
> > > The two partitions are on different disks, so each one should be
> > > isolated from the I/O perspective (do they share a single
> > > controller?).
> > 
> > No, sda/sdb use separated controller. 
> >  
> > >  The only interaction I see happens at the VM level,
> > > since changing performance of any of the two changes the rate at which
> > > pages can be cleaned.
> > > 
> > > Corrado
> > > >
> > > >
> > > > vmstat average output per 10 seconds, without ccache
> > > > procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu-----
> > > >  r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
> > > > 26.8 0.5 0.0 63930192.3 9677.0 96544.9 0.0 0.0 2486.9 337.9 17729.9 4496.4 17.5 2.5 79.8 0.2 0.0
> > > >
> > > > vmstat average output per 10 seconds, with ccache
> > > > procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu-----
> > > >  r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
> > > > 2.4 40.7 0.0 64316231.0 17260.6 119533.8 0.0 0.0 2477.6 1493.1 8606.4 3565.2 2.5 1.1 83.0 13.5 0.0
> > > >
> > > >
> > > >>
> > > >> Jan,
> > > >> can you send a patch with similar change for ext3? So we can do more tests.
> Hi Jan,
> can you send a patch with both ext3 and ext4 changes? Our test shows
> your patch has positive effect, but need confirm with the ext3 change.
  Sure. Patches for both ext3 & ext4 are attached. Sorry, it took me a
while to get to this.

									Honza
-- 
Jan Kara <jack@...e.cz>
SUSE Labs, CR

View attachment "0001-jbd2-Refine-commit-writeout-logic.patch" of type "text/x-patch" (11044 bytes)

View attachment "0002-jbd-Refine-commit-writeout-logic.patch" of type "text/x-patch" (9844 bytes)

Powered by blists - more mailing lists