[KLUG Members] Two questions.

[Adam Tauno Williams]

>2.  Does anyone know of any hard evidence (I've found some heresay and
>conjecture, those are kinds of evidence) that linux boxes read and write
>slowly to NFS mounts exported from BSD machines.  I've got a few linux 
>machines writing to an OpenBSD server, exporting it's RAID to those servers.
>Writes and reads to the NFS drive are slow. Far slower than they should be. 
>The network environment is switched 100Mbit, so we should get at least 6
>MB/sec, right? Heck, I can get 8 or 9 on my hub using SAMBA, for goodness'
>sakes. Anyway, we are languishing around 2 MB/sec on some, whereas on others,
>we have bumped it up a bit.  So far, the best block size is 8192 bytes. 
>standard on our boxes is 2048 (yeah, that's pretty small). 

Do you have FreeBSD NFS clients that are performing better?

How many Linux clients do you have?  What is the proportion of clients to NFS
server threads?

What is the network I/O buffer size?  By default this is way to small to get
good performance on large network transactions.

I've got a handful of slides about tuning the network buffers and NFS in a
presentation,  I'll stuff it up on the KLUG FTP site.

ftp://kalamazoolinux.org/pub/pdf/PerfTune2001.pdf

Systems and Network Administrator
Morrison Industries
1825 Monroe Ave NW
Grand Rapids, MI. 49505