Re: Maximum performance for caching

On Mon, Sep 06, 2004 at 10:42:56AM +0900, Batara Kesuma wrote:
> Yes, more than 80% of the requests are static contents. TUX is amazing,
> but the problem is the NFS server :( 

May be able to tune the NFS server a bit to improve performance. IF the
NFS server is in the same physicaly building as the front-end machines,
you likely don't need TCP-based NFS; switch to UDP. I ran NFS-based
webservers for a couple of years, and the performance benefit from using
UDP is significant. You can also increase the size of the NFS buffer
size. Ultimately, however, NFS is rarely as fast a local disk (a
possible exception to this is using a NetApp NFS server; NetApp swears
up and down that their RAID4/WAFL systems are on-par with local disk for
access times; it's what I ran when I was using NFS for my webservers).
If you need a central disk array, you could look into a SAN solution
(EMC, etc) if you have the money. Most people don't. =)

Another possibility is that cachefs patches were recently produced (see:
http://lwn.net/Articles/99597/) which *could* improve your NFS
situation. Note that I stress *could*; that code has *not* been merged
into the kernel tree, and likely still has a few lingering bugs.
Additionally, I've not used it, so can't advise as to what gotchas may
be lurking around the corner. YMMV. =)

--n

-- 
We need to fold the monkey.