Anyone have information on latest LSU beowulf?

[Craig Tierney]

On Tue, Oct 08, 2002 at 01:35:27PM +0100, Daniel Kidger wrote:
> 
> On Mon, 23 Sep 2002, Craig Tierney wrote:
> > http://www.phys.lsu.edu/faculty/tohline/capital/beowulf.html
> > 
> > Their HPL result is 2.2 Tflops!  Very impressive.
> > 
> > (lines deleted) 
> > What HPL settings did they use to achieve their result?
> 
> I think that many people would be interested in their settings for xhpl,
> in particular what percentage peak did they manage for a single CPU run ?
> 
> A 1.8GHz P4 has a theoretical peak of 3.6GFlops/s, but so far I have only
> seen figures of around 60% of this for linpack. Compare this with 75%+ for
> Alpha nodes (and of course 95%+ for vector processors).
> 
> So, in terms of single processor performance:
> 
>    Which compiler did they use ?  (icc version 7 perhaps)
>    And which compiler options ?
>    Did they use mkl or Atlas for the BLAS ?

Probably Atlas.  I tested both and got about 70-80% more speed
out of Atlas compiled with gcc-2.91 (recommended).

>    What value of NB did they settle on ? (80 and 160 seem common choices)
>    any other non-default values in HPL.dat ?

Why are 80 and 160 common choices?  I do know that they used 160
for their run.  I also retested my setup at 160 and it is much
slower than 64.  I was told by someone at UTK that the size of
NB should be a multiple of the L1 cache and that double is good.
So NB = sqrt(8kb * 1024/8)=32 for P4 Xeon.  I tried 64 and that has 
been the best for a single node run.  

The one thing I have not yet determined is that maybe the larger number
is more effective when you start running on large node counts.
I cycled through the NB values from 32 to 160 on 64 processors
and NB=64 was still the best.  

I wonder if having more memory (1 GB vs. 2 GB per node) could
drastically improve scaling.  Anyone know?

Craig

-- 
Craig Tierney (ctierney at hpti.com)


> 
> 
> 
> 
> Yours,
> Daniel.
> 
> --------------------------------------------------------------
> Dr. Dan Kidger, Quadrics Ltd.      daniel.kidger at quadrics.com
> One Bridewell St., Bristol, BS1 2AA, UK         0117 915 5505
> ----------------------- www.quadrics.com --------------------