🚀 go-pugleaf

WEILL Philippe <philippe.weill@aero.jussieu.fr> writes:

>When I use my OSF1 box as a server for my linux (i586) box 
>,performance is pretty bad

Standard problems with Digital UNIX NFS servers vs Linux clients include:

1) Adherence to the NFS V2 spec.
   Digital's systems don't respond to write requests until the data and metadata
   are on the disk.  HP and SGI don't do this and can lose data if the
   server crashes while NFS writes are being done.

2) "Write gathering" in the server.
   This is trying to get several writes to do them all at once.  This is great
   on Unix where the NFS clients tend to send several writes at once.  I think
   Linux doesn't do that, and a sleep delay in write gathering does you in.
   Try patching the kernel variable nfs_ufs_lbolt to be zero.  This will
   quicken the write replies and not have too great an impact on other
   Unix clients.

3) Buy a presto board for the server.
   This will have a big improvement.  It will also speed up local directory
   manipulation on the server (like speed up tar extraction, etc.)

4) Implement NFS V3 on Linux.  :-)
   V3 supports safe asynchronous writes and fixes the V2 write performance
   problem the right way.  Over FDDI, we can get V3 writes at 6 MB/sec
   without Presto (but with striped disks).

Ah - yes, 8KB transfers are the largest we support.
-- 
Eric (Ric) Werme         |  werme@zk3.dec.com
Digital Equipment Corp.  |  This space intentionally left blank.