[Beowulf] Strange Behavior

Dan Kidger daniel.kidger at quadrics.com
Sat Mar 13 17:00:40 EST 2004


On Friday 12 March 2004 4:43 pm, Lars Henriksen wrote:
> On Fri, 2004-03-12 at 13:51, Mathias Brito wrote:
> > I'm benchmarking my 16 nodes cluster with HPL and I
> > obtain a estrange result, different of all I ever seen
> > before. When I send more data with a big N, the
> > performance is worse than with small values of N. I
> > used N=5000 with NB=20 and the performance was 3.3GB,
> > when I send N=10000 with NB=20 i get only 2.1GB. I
> > don't liked the result, the nodes are athlon xp 1600+
> > with 512MB RAM, and I think the cluster very slow.
> > Someone had the same problem and could help me?
>
> Please correct me anybody, if im wrong:
> It seems to me, that the best results are acheived with approx 85-90%
> memory utilization (leaving something to the rest of the system).
>
> (16*512*1024*1024/8)^0.5 ~= 30200, that would close to the best N value


Your target should be say 75% of theoretical peak performance 
   0.75 * 16nodes  * 1 cpupernode * 1.4Ghz * 1 floppertick 
   = 16.8 Gflops/s

So figures like '3.1' Gflops/s (14% peak) are much lower than what you should 
be achieving (Only vendors like IBM post figures on the top500 with %peak 
figures as low as this (Nov2003) )

Linpack figures are dominated by the choice of maths library - you do not say 
which one you are using (MKL, libgoto, Atlas, ACML) ?


> isn't Nb=20 very low? I currently use arround 145 for P4 cpu's

Remember choice of NB depends on which maths library you use rather than 
simply on the platform - but in general the best values lie between 80 to 
256; 20x20 is far too small for a matrix multiply.


Daniel.

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



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list