[Top] [All Lists]

Re: [MMTests] IO metadata on XFS

To: Mel Gorman <mgorman@xxxxxxx>
Subject: Re: [MMTests] IO metadata on XFS
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 2 Jul 2012 09:54:58 +1000
Cc: linux-mm@xxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx, linux-fsdevel@xxxxxxxxxxxxxxx, xfs@xxxxxxxxxxx
In-reply-to: <20120629112505.GF14154@xxxxxxx>
References: <20120620113252.GE4011@xxxxxxx> <20120629111932.GA14154@xxxxxxx> <20120629112505.GF14154@xxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Fri, Jun 29, 2012 at 12:25:06PM +0100, Mel Gorman wrote:
> Configuration:        global-dhp__io-metadata-xfs
> Benchmarks:   dbench3, fsmark-single, fsmark-threaded
> Summary
> =======
> Most of the figures look good and in general there has been consistent good
> performance from XFS. However, fsmark-single is showing a severe performance
> dip in a few cases somewhere between 3.1 and 3.4. fs-mark running a single
> thread took a particularly bad dive in 3.4 for two machines that is worth
> examining closer.

That will be caused by the fact we changed all the metadata updates
to be logged, which means a transaction every time .dirty_inode is

This should mostly go away when XFS is converted to use .update_time
rather than .dirty_inode to only issue transactions when the VFS
updates the atime rather than every .dirty_inode call...

> Unfortunately it is harder to easy conclusions as the
> gains/losses are not consistent between machines which may be related to
> the available number of CPU threads.

It increases the CPU overhead (dirty_inode can be called up to 4
times per write(2) call, IIRC), so with limited numbers of
threads/limited CPU power it will result in lower performance. Where
you have lots of CPU power, there will be little difference in


Dave Chinner

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