On Mon, 2002-07-15 at 14:57, Austin Gonyou wrote:
>
> In the following directory in proc, /proc/sys/fs/xfs/, exists the
> following tuneables, refcache_purge refcache_size . What, if any, can I
> gain by tweaking these on a quad processor P4 Xeon server? I'm using
> 2.4.19-rc1-aa2 on this box and pushing ~900GB oracle DB. I've got 8GB
> ram also.
Those are for the NFS refcache, which holds references on files written
via NFS, to keep pre-allocation in-place for a while. To quote Steve,
"This prevents xfs from constantly freeing this space and then
reallocating it on the next write and is a very major performance win in
NFS."
So, the size & purge sysctl values are how big this cache is, and how
many entries are purged out every time the filesystem syncs (i.e. how
fast it drains).
If you have a lot of writes coming in at once over NFS making the cache
larger might help. In general you'd probably scale the purge size with
the overall size.
Since it holds things open with pre-allocation, you'll see more disk
usage while the file is in the NFS refcache. If you have quotas, this
could impact user quotas. If you have a lot of dirty entries in the
refcache, umount might take longer, as well.
If you find anything interesting while testing, let us know. :)
-Eric
|