Re: Performance testing of C/R

From: Brent Gorda (bgorda_at_lbl.gov)
Date: Wed Feb 19 2003 - 15:33:07 PST


I think they have 24 nodes on that system.  Lots of storage targets
too.  Whether they'd take an interest in this, we'd have to investigate.

Brent


On Wed, 2003-02-19 at 15:41, Eric Roman wrote:
> 
> It sounds like a good idea.
> 
> What I'm saying is that I don't know if we're capable of generating data
> that quickly.  We might have to fiddle with it for a week or two to
> make it fast.
> 
> How many nodes are on GUPFS?  If it really is a parallel FS, we might be
> able to saturate with multiple nodes...
> 
>  - E
> 
> On Wed, Feb 19, 2003 at 03:03:38PM -0800, Brent Gorda wrote:
> > Mike is seeing 1.4 GByte/sec on the Yottayotta box.  Insert lots of
> > disclaimers (like no file system ;-), but that is a pretty interesting
> > number.  I think some straightforward (ie student) work on this front
> > might net you all a good paper or some excellent PR.
> > 
> > Brent
> > 
> > On Wed, 2003-02-19 at 15:00, Eric Roman wrote:
> > > On Wed, Feb 19, 2003 at 02:15:57PM -0800, Brent Gorda wrote:
> > > > 
> > > > In talking with Mike, I wondered if it might not be a fun exercise for a
> > > > student this summer to put the C/R code up on the GUPFS cluster and get
> > > > some timings on the various hardware they are using over there.
> > > > 
> > > > Thoughts?
> > > 
> > > We've got a lot of room to speed up C/R.  It's not very fast at the moment.
> > > First on the list would be implementing our own write buffers...  That should
> > > speed us up quite a bit.
> > > 
> > > More performance testing even on our own cluster migt not be a bad idea!
> > > 
> > >  - E
> > > 
> > > > 
> > > > Brent
> > > > 
> > > 
> > > -- 
> > > Eric Roman                       Computational Research Division
> > > 510-486-6420                     Lawrence Berkeley National Laboratory
> > 
> 
> -- 
> Eric Roman                       Computational Research Division
> 510-486-6420                     Lawrence Berkeley National Laboratory