On April 26, Bruce Guenter wrote:
> So far, XFS with an external journal has by far the best delivery rate.
> However, as the filesystem fills and delivery concurrency increases, the
> time required to list, read and delete the delivered files slows down to
> unacceptable levels. Is there anything I can do, settings to modify,
> patches to try, to improve this behavior?
Some generic suggestions:
- more log buffers than fewer (try 8)
- It looks like you're doing a lot of metadata intense operations, so
you should try a larger log buffer size which are available in
version 2 logs. I wrote version 2 logs, along with some other
changes, to speed up metadata intense benchmarks. I think 128mb
log buffers work on linux.
Stay away from log striping. The last I heard, it still doesn't work.
- You didn't say what your filesystem size is. There is an ugly hack
we put in to kee inode numbers < 2**32. This changes inode and
data placement (inodes are kept in the lower AGs, user data is
spread around in the upper AGs) and brings on the full-filesystem
case earlier. I think thats what you hit.