Scyld performance problems

[Karen Keadle-Calvert]

Tom,

If you are running the linpack that came with your Scyld distribution, 
it must be optimized for the cluster you are running on or the results 
are not very meaningful.  /usr/bin/linpack is just a script that runs 
xhpl for you automatically by creating the input file for you.
Please refer to the file /usr/share/doc/hpl-1.0/TUNING to get ideas on 
optimizing the input file.  The easiest/quickest change to make is to 
increase the problem size.  By default, the problem size is 3000. 
Beginning with release 27cz-9, you can change the size on the command 
line by typing: linpack -n size.  Otherwise, you'll have to modify the 
/usr/bin/linpack script to change the problem size.  Just note that if 
you make the problem size too large the nodes could run out of memory 
and the test will fail (you'll see a 'connection refused' message when 
this happens).

Hope that helps.

Regards,

Karen Calvert
Scyld Computing

Tom A Krewson wrote:
> I am having a performance problem with my Scyld Beowulf. I have been
> testing with 5 machines on a 100-megabit switch while using the Linpack
> benchmark. With 6 machines (1GHZ server, 5-333 MHZ clients) the results of
> the linpack benchmark is .1 Gflops. With just one machine (1GHZ server) I
> get a .5 Gflops (4.670e-01) result. I tried using 9 computers and the
> performance is even less at .09 Gflops. I have tried 3 different switches,
> and the cheapest (I assume a cut through switch) gives the best
> performance. 
> 
> I have downloaded a parallelized pi calculation program the reverse true -
> the more computers in the cluster, the faster the calculation.  
> 
> Am I reading the results incorrectly? What might cause this poor
> performance?
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf