Hi,
I wonder if the differences here are to do with the server / network,
rather than NFS per se. We had 6 processor Sun NFS server with a
pretty clean network, for example. The tests we did were with an
Alpha and another Sun as clients, also.
Best,
Matthew
On Thu, 06 Jan 2005 10:00:40 -0600, Ignacio Valdes <[log in to unmask]> wrote:
> Our setup is Suns with local apps and a Linux server that holds all
> the data. We had a big increase in speed after changing to Gigabit
> Ethernet. I/O over our Intranet is still not ideal however. We suspect
> the overhead from NFS. We speculate that thin clients are the way to
> go in the future to improve performance by avoiding the NFS overhead.
> Does anyone have opinions on this?
>
> -- IV
>
> On Thu, 6 Jan 2005 07:11:10 -0800
> Matthew Brett <[log in to unmask]> wrote:
> > Hi,
> >
> >> That depends on how many analysis is done and the size of the fMRI
> >> dataset. We tried running 3 analysis on a fast dual-processor
> >>machine
> >> recently and find that there is a bottleneck on NFS @1Gbps.
> >
> > That's interesting; we didn't notice much extra slowdown for 6
> > processor machine running 6 realignment jobs on a 100 Mbs network.
> > This was an Alpha machine with processors running at rough
> >equivalent
> > of 2 GHz Pentiums I suppose. Did you do any direct timings by any
> > chance, comparing NFS to local hard disk? Did you saturate the NFS
> > bandwidth?
> >
> > Best,
> >
> > Matthew
>
>
|