[molpro-user] Shared memory problem using RPMs

[Nick Wilson]

Scott,

molpro checks that the machine returned by gethostname is in the node
list. There is a --mpirun-nolocal flag to molpro but I'm not sure if
this is the right thing to disable this. If you can't get gethostname to
return the same name that pbs uses then alternatively you can either use
the -N flag to specify the nodelist, or use sed to edit $PBS_NODEFILE in
your job script.

Opteron build instructions are here:
http://www.molpro.net/supported/opteron.php

The portland compiler (5.1 or 5.2) works fully in parallel. 5.2 is
quicker than 5.1. If you're using version 5.2 of the portland compiler
then make sure the same flags are being used when building global arrays
and within molpro.

The intel premier support page has just released version 8.1 of the
intel compiler which has the EM64 technology which is analagous to
x86_64 so can produce 64bit code which runs on opteron. I've produced a
working serial executable but configure doesn't work yet and you have to
tweak the CONFIG file. I haven't yet tried to build global arrays with
this compiler.

The pathscale compiler http://www.pathscale.com also works in serial but
not yet in parallel. I will backport the configuration for this onto
2002.6 soon.

Some benchmark numbers here:
http://www.molpro.net/benchmark/bench.cgi?fields=wall&fields=total&fields=xhost&nompp=1&nothroughput=1

Nick Wilson

Scott Yockel wrote:
> Molpro users,
> 
> I am having a shared memory problem on my dual-proc Opteron based cluster under
> linux.  Currently I'm running latest 32-bit Alhlon molpro 2002.6 mpp builds.  It
> runs fine in dual proc when running interactivly submitting strait from the
> command line.  However, when submitting through a batch system (OpenPBS) then I
> get the following output:
> 
> Nolocal= 1
> Parlib= 0
> Local machine n06.priv0.unt.edu not in node list!
> 
> What I assume is different is the tty control from the queuing system has now
> lead molpro to try to look for distributed memory and other nodes, which we only
> want it to run a local 2proc SMP-type job.  If the "Nolocal=1" is a boolean type
> (true/false) switch, can it be turned off, and how?  Or is there any other type
> of environment variable that I'm not aware of that needs to be set to fix this
> problem?  Also, this system came prebuild with quite a bit of mpi and pvm type
> of libraries intended for distributed memory.  Is is possible that they share
> something in common with molpro that might foul up running 2proc SMP-type jobs
> on these operons locally ?  
> 
> Also, has there been any completely successful 64-bit compiles of Molpro2002.6
> on a Opteron based cluster, and if so are there any makefiles for this?
> 
> Thanks so much for the help,
> 
> Scott Yockel
> University of North Texas
> Department of Chemistry
> New CHEM building
> Room 262C
> 
> 
> 
>