[Beowulf] 64bit comparisons

[Hujsak, Jonathan T (US SSA)]

Hi!

 

We're looking at implementing a large G5 cluster here at BAE Systems.

 

Have you gained any new 'lessons learned' since the communication 

below? Can you recommend a good version of MPI to use for these?

We've been looking at MPICH, MPIPro and also the Apple xgrid...

 

Thanks!

 

Jonathan Hujsak

BAE Systems

San Diego

 

Bill Broadley bill at cse.ucdavis.edu
<mailto:beowulf%40beowulf.org?Subject=%5BBeowulf%5D%2064bit%20comparison
s&In-Reply-To=200405141644.i4EGi1Aq023213%40marvin.ibest.uidaho.edu> 
Fri May 14 11:48:21 PDT 2004 

*	Previous message: [Beowulf] 64bit comparisons
<http://www.beowulf.org/pipermail/beowulf/2004-May/016707.html> 
*	Next message: [Beowulf] 64bit comparisons
<http://www.beowulf.org/pipermail/beowulf/2004-May/016724.html> 
*	Messages sorted by: [ date ]
<http://www.beowulf.org/pipermail/beowulf/2004-May/date.html#16711>  [
thread ]
<http://www.beowulf.org/pipermail/beowulf/2004-May/thread.html#16711>  [
subject ]
<http://www.beowulf.org/pipermail/beowulf/2004-May/subject.html#16711>
[ author ]
<http://www.beowulf.org/pipermail/beowulf/2004-May/author.html#16711>  

  _____  

On Fri, May 14, 2004 at 09:44:01AM -0700, Robert B Heckendorn wrote:
> One of the options we are strongly considering for our next cluster is
> going with Apple X-servers.  There performance is purported to be good
 
Careful to benchmark both processors at the same time if that is your
intended usage pattern.  Are the dual-g5's shipping yet?  Last I heard
yield problems were resulting in only uniprocessor shipments.  My main
concern that despite the marketing blurb of 2 10GB/sec CPU interfaces
or similar that there is a shared 6.4 GB/sec memory bus.
 
> and their power consumption is small.
 
Has anyone measured a dual g5 xserv with a kill-a-watt or similar?
 
> Can people comment on any comparisons betwee Apple and (Athlon64
> or Opteron)?
 
Personally I've had problems, I need to spend more time resolving them,
things like:
*       Need to tweak /etc/rc to allow Mpich to use shared memory
*       Latency between two mpich processes on the same node is 10-20
times the 
    linux latency.  I've yet to try LAM.
*   Differences in semaphores requires a rewrite for some linux code I
had
*   Difference in the IBM fortran compiler required a rewrite compared
to code
    that ran on Intel's, portland group's, and GNU's fortran compiler.

 
Given all that I'm still interested to see what the G5 is good at and
under
what workloads the G5 wins perf/price or perf/watt.
 
-- 
Bill Broadley
Computational Science and Engineering
UC Davis

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.scyld.com/pipermail/beowulf/attachments/20041008/19c8e119/attachment.html