[Top] [All Lists]

Re: observed significant performance improvement using "delaylog" in

To: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Subject: Re: observed significant performance improvement using "delaylog" in
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Sat, 14 Aug 2010 00:13:44 +1000
Cc: Michael Monnerie <michael.monnerie@xxxxxxxxxxxxxxxxxxx>, Stan Hoeppner <stan@xxxxxxxxxxxxxxxxx>, xfs@xxxxxxxxxxx
In-reply-to: <20100813122907.GA14650@xxxxxxxxxxxxx>
References: <201008121346.30760.eye.of.the.8eholder@xxxxxxxxx> <i41q43$6td$1@xxxxxxxxxxxxxxx> <4C6516CA.2010602@xxxxxxxxxxxxxxxxx> <201008131235.45050@xxxxxx> <20100813122907.GA14650@xxxxxxxxxxxxx>
User-agent: Mutt/1.5.20 (2009-06-14)
On Fri, Aug 13, 2010 at 08:29:07AM -0400, Christoph Hellwig wrote:
> On Fri, Aug 13, 2010 at 12:35:44PM +0200, Michael Monnerie wrote:
> > On Freitag, 13. August 2010 Stan Hoeppner wrote:
> > > Some benchmark results maybe worth a look:
> > > http://btrfs.boxacle.net/repository/raid/2.6.35-rc5/2.6.35-rc5/
> >  
> > Thanks - it would have been great to see xfs with delaylog in that 
> > comparison, but the graphs are very very nice.
> > 
> > XFS seems performing better the more threads there are, just in "large 
> > file random reads" it's the slowest - why this?
> Any idea who is doing these runs?

IIRC the tests are run by someone from IBM, but I cannot remember
who it is.

> Once we figure out what that large
> file random reads loads is I'm sure we could fix it soon.

>From http://btrfs.boxacle.net/:

Random Reads (raid, single-disk) 
        Start with 1024 files. 
                100 MB files on the raid system.
                35 MB files on the single-disk system.
        Each thread reads a fixed amount of data from a random location in one 
file using 4 kB reads. 
                5 MB reads on the raid system.
                1 MB reads on the single-disk system.

So it's not a small random read workload (100GB data set), so the
files on XFS are probably more spread out over multiple AGs
and hence further apart than other filesystems. Hence a greater
average seek distance, hence it slower throughput....

> And asking
> him/her to add -o delaylog would also be good.

Yes, that would be an interesting comparison...



Dave Chinner

<Prev in Thread] Current Thread [Next in Thread>