[ILUG] NFS throughput drops with SCSI activity

[Niall O Broin]

On Thu, Apr 26, 2001 at 12:48:59PM +0100, Kenn Humborg wrote:

> 
> Machine:    HP Vectra PPro 200MHz 
>             dual-proc board with 1 CPU
> Kernel:     2.2.18 uniprocessor
> Ethernet:   Netgear FA311 100BaseT
>             Netgear-supplied driver
> SCSI:       Adaptec AIC-7880U on-board
>             aic7xxx driver
> DISK:       Seagate ST32171N
>
> Reading the image and dumping it to /dev/null gives me
> 2MB/s average across the network.  (In fact, it peaks
> around 4.5MB/s, but is very bursty - don't know why, but
> don't really care, since the CPU can only decompress at
> about 2MB/s average anyway.)
> 
> dd-ing from /dev/zero to the SCSI disk gives about 6MB/s,
> which is respectable enough.  Given an approx compression
> ratio of 3:1 (as seen in practice) this matches nicely with
> the network and CPU throughput.
> 
> However, while writing to the disk, NFS throughput drops
> to about 0.5MB/s and I get "server timed out" & "server OK"
> messages a lot.  Sometimes NFS practically halts.  (Note
> that this happens even when the NFS and disk commands are
> not in the same pipeline, so it's that the pipes are getting
> full.)
> 
> What could be causing the slow-down and how might I fix it?

This is a tasty one, Kenn ! The only thing that jumps to mind is scheduling
/ processor load but I have to admit, it sounds unlikely. I considered
gzip's consumption of CPU which will be high, but you say you've done the
whole thing with dumping it to /dev/null so that should let out gzip. Have
you tried watching top while writing to /dev/null, and while writing to the
disk ? Perhaps the aic7xxx driver is greedy ? Have you tried running Windows
and using Ghost :-)




Niall