you should compile with OpenMP support (which also means,
that you have to ditch the crappy default PGI programming environment
and use gcc/gfortran instead).
for plain quickstep calculations there should be a
significant performance increase when running with
4 MPI tasks per node and 8 threads per MPI task.
due to the massively reduced need for data
replication, you should not run out of memory.
since interlagos CPUs do some kind of improved
hyperthreading and two tasks will have to share
a floating point unit, i would expect that you should
get a large improvement already by just using
only half the cores when using plain MPI only.
make sure that you use the proper "distance"
flag to aprun, however, since you need to make
sure you have exactly one MPI task per
"bulldozer" subunit.
i would talk to the user support specialists
of the machine, to make sure you get it right.
cheers,
axel.