On Wed, Apr 06, 2011 at 02:33:56PM -0400, Christoph Hellwig wrote:
> The patch looks correct, but I wonder what the performance implications
> of flushing the AIL every 30 seconds are. Did you benchmark the
> patchset to see what it does to performance numbers for metadata
> intensive workloads?
In my measurements it's made no difference.
If the workload is being sustained for more than a few seconds, then
memory reclaim typically becomes active and the push from there
becomes the dominating factor. This effectively makes the syncd work
trigger a no-op.
If the push from the syncd work does trigger work to be done, it's
generally because the filesystem has gone to idle or only has a low
level of modifications being done. In either case, it doesn't affect
performance, especially with delaylog pinning the working set of
metadata in memory...
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|