fluent on linux

Patrick Lesher plesher at ibb.gatech.edu
Tue Oct 17 11:48:35 PDT 2000


We set up a small ( 8 -10 cpu ) cluster of general machines just for
testing fluent and a proof-of-concept design.  The cluster was a set of
BP-6's, 400-600mhz CPU's, 256 megs of ram, and switched 100mbit.

One job went from 8 hours ( PII 400 and 128 megs of RAM ) to between 20
and 30 minutes on the cluster.  Granted most of the speed up was the extra
ram, we were still really pleased.  We also found ( understand our lack of
Fluent experience ) that using non-power-of-2 sets of CPU's actually hurt
our performance.  Also, running it on our brand new Sun Ultra 10's with
the same number of CPU's gave slower results then our PC cluster.

It has been a while, but our general feel was that performance was great
on the simple cluster.  Tuning the fluent code/job gave the best %
increase.  Give it as much RAM as you can.  And the network didn't see to
be that much of a bottleneck.  Granted these will change per job and per
situation.

Patrick

On Tue, 17 Oct 2000, Ivars Drikis wrote:

> Hi,
> 
>   Some days ago there where question about experience of running
> Fluent on boewulf cluster. At present moment we are still running
> serial versions at SGI, but we have found following information
> about this possibility at May 2000.
> 
>   I first, HP-Fluent-Myricom Fluent on HP Clusters. After filling
> of prommising web form at Fluent homepage, we get with e-mail 
> list of components as well as e-mails from Fluent, Myricom
> and SGI sales departments (or local distributors). The cluster 
> specifications includes 4 or 16 HP XL550 dual systems, Myrinet 
> or Ethernet, WinNT or RedHat 6.0. No information concerning 
> benchmarks, parallel efficiency and price. 
> 
>   Next we connect our local Fluent distributors in Finland, in
> particular Björn Jernström, bjorn at processflow.fi. We get information
> that they runs three PIII machines interconnected by a 100MHz Fast 
> Ethernet, and "the parallel efficiency is typically close to 100%".
> So, maybe good idea is ask them gain.
>   
> ---------------------------------------------------------------
> Dr. Phys. Ivars Drikis       Faculty of Physics and Mathematics
> 				    University of Latvia
> phone (ofice 1): 371-7-615712        Zellu str. 8, F341 
> phone (ofice 2): 371-2-945830       Riga, Latvia, LV-1002
> fax            : 371-7-901214     e-mail:  drikis at lanet.lv
> 
> 
> 
> 
> 
> 
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf
> 





More information about the Beowulf mailing list