Re: [Fwd: [Fwd: xfs write speed regression 2.6.18.1 to 2.6.19.1]]

[David Chinner <dgc@sgi.com>]

On Fri, Jan 05, 2007 at 04:19:48PM -0600, Dave Lloyd wrote:
> From a co-worker. Anyone know what might have changed this between
> 2.6.18 and 2.6.19 when the issue first appeared?

IIRC< a bunch of changes went into the generic buffered I/O
path to fix deadlocks on writes if we take a page fault during
the copyin. That caused a performance regression for buffered
I/O of around that sort of figure, and the regression is slowly
being fixed up as per:

> under 2.6.20-rc3 the speeds have gone back up some, but they are 10%
> slower than 2.6.18.

So I don't think this is an XFS problem as such. Still, I will try
to do some local tests to check it out.

> and the allocations, as shown by the sequential writes (attached) are
> random.

????

> If I went all the way out to the inside tracks, you would be at about
> 490MB/Sec.
> 
> Something changed. 2.6.19 was unstable, with XFS panics on a regular basis.

Got any stack traces?

> 2.6.19.1 has not had an error yet.. (knock head on wall repeatedly).

We didn't push any changes into 2.6.19.1, so that implies bugs in
the generic code, not XFS....

Cheers,

Dave.
-- 
Dave Chinner
Principal Engineer
SGI Australian Software Group