[Beowulf] 64bit comparisons

Robert Latham robl at mcs.anl.gov
Mon May 17 07:08:13 PDT 2004


On Fri, May 14, 2004 at 11:48:21AM -0700, Bill Broadley wrote:
> 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.	

This might sound like a troll, but I'm serious: why not run linux on
the G5 Xserves?  That would address point 1, 2, and 3 and you would
have a much better underlying kernel.  The Darwin kernel is a terrible
performance dog, and linux outperforms it in context switches, file
system access, page faults and just about every other kernel-specific
metric by an order of magnatude.  It's a cluster, so you're not using
the OS X GUI.  You've probably already got a cluster management
infrastrucutre for your linux clusters, so you won't have an oddball
Darwin cluster screwing things up.  

I don't know how possible it is to get 3rd party software for
linux-ppc.  Aside from that one issue, linux on G5 clusters seems to
make a lot of sense from a portability, performance, and
administration perspective.
 
==rob

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Labs, IL USA                B29D F333 664A 4280 315B



More information about the Beowulf mailing list