[Beowulf] Gaussian in parallel

[Reuti]

Hi,

> Interesting.  I was under the impression that SDSC had
> a linux cluster that ran g03 in parallel.  I'm not exactly sure
> of their relationship, I guess not all SDSC systems are
> part of the NCSA.
> 
> We do have it running on two clusters here, it was not
> trivial.  The hardest part was using a supported platform
> and integrating that into our environment.  We ended up
> using Redhat 8 for it, not willingly.  I chatted with the folks
> at Gaussian and different people gave me contradictory
> information as to whether it worked on more modern
> Redhat variants (RHEL, Centos, WBEL and the like).
> 
> I tested it on Centos and ran  into problems.
> 
> On RH8, we the trick was to make sure we were calling
> the linda binaries and using scratch files that were accessible

What was the problem in detail? The only thing I changed was in linda_rsh the 
definition of /usr/bin/rsh to rsh to use the SGE (Sun Grid Engine) wrapper to 
control the slave tasks.

> across the cluster.

This is not necessary. You need identical named scratch directories on of all 
the nodes, which we create with SGE during startup of the job and this is 
working (and remove it after the job).
 
> Also, g03 (according to Gaussian) is still 32-bit in Linux
> environments and may be memory bound as a result.

You can compile it since C.02 also with AMD64 as 64 bit native and have big 
scratch files without any 16GB limit.

Cheers - Reuti