Re: [FLASH-USERS] speedup

From: Anshu Dubey <dubey@tagore.uchicago.edu>
Date: Fri Dec 03 2004 - 09:02:07 CST

In order to get speedup with any parallel code, you have to have enough
work to do per processor, because there are always latency issues. To
really judge how good or bad your cluster is, I suggest you run the
scaling_test. It is a setup in the distribution of the code. It keeps
increasing the size of the problem as you increase the number of
processors. If your machine has good communication characteristics, and
you plot number of procs along X and execution time along Y, you should
get a straight line parallel to X axis.

On Fri, 3 Dec 2004 M.Krause@lsw.uni-heidelberg.de wrote:

> Hello again,
>
> I test flash to run on a linux cluster with 100Mbit ethernet connections.
> Running the sedov test, I do not get any speedup increasing the number of
> processors. I have run the mpich (1) performance test. It indicates the
> maximum data transfer rate (100Mbits/s) from message sizes of about 10,000
> Byte and bigger.
>
> Does flash simply need faster connection hardware for speedup?
> Does anybody run flash on a similar cluster, gets speedup and would like
> to share the trick?
>
> Thanks for any comments & regards
> Martin
>

-- 
Anshu Dubey
Astronomy & Astrophysics
University of Chicago
5640 S. Ellis Ave.  Chicago IL 60637
Tel : (773) 834 2999
Fax : (773) 834 3230
Received on Fri Dec 3 09:11:34 2004

This archive was generated by hypermail 2.1.8 : Thu Aug 31 2006 - 21:20:48 CDT