[gmx-users] How can we have better performances on an IBM JS20- Myrinet-Linux cluster?

[Pietro Lopriore]

Hi everybody, sorry to post so many questions, but I hope that someone can
help us!

Our environment is the following:

JS20 (2processors per node, 28 nodes) with Linux Suse 9 kernel
2.6.5-7.97-pseries64
Myrinet, GM 2.0.16-2,
MPICH-GM 1.2.6.14,
FFTW 2.1.5
gcc 3.3.3-43.24
gcc-64bit-9-200407011606

------PROBLEM---------
We're actually working with the MPI machine file without inter-node
communication. In fact GROMACS seems very very slow when listing all the
nodes' processors.
In order to maximize the performances or at least to optimize the usage of
the software (it scales very well until up to 8-16 CPUs, as the benchmarks
suggest) we'd like to have a mixed shared memory - Myrinet (MIPCH-GM)
communication.
In order to have better performances (as suggested by GROMACS' site) we
compiled it in simple precision but in this case we had to force the
Altivec "disable" flag to have it correctly compiled.
When we try to use a different machine file with all the single nodes
listed we had serious performance problems.
------QUESTIONS--------
Is it a problem connected to the MPICH-GM compilation (how can we compile
with both SMP support and inter-node communication)?
Could we obtain better performances compiling all with IBM compilers and
installing Scientific Libraries?
Do you think that using the Fortran compiler could produce better results
for some parts of the code?
How can we use the Altivec (VMX) libraries? Is that a gcc version issue
(altivec-support) ?

Sorry for all these questions.
Any help will be greately appreciated. Many many thanks!

Pietro

_________________________________________________

Pietro Lopriore
EMEA South Region Technical Sales Support
IBM Healthcare and Life Sciences Industry