Parallel communication - SHMEM / ARMCI

Hello!

I'm currently implementing an algorithm which could profit from running
on a massiv parallel system but has to communicate efficiently. So far,
it boils down to these methods:

1) MPI-1: not very efficient in our case

2) MPI-2 (one-sided communication): could be interesting, but so far
lacking support on our machines

3) SHMEM: Seems nice, but lock-in into Crays and SGIs - which is not so
bad in our case, because we do have one of these ...

4) ARMCI: (http://www.emsl.pnl.gov/docs/parsoft/armci/) benchmarks show
extremely good performance, but not available per se for
Fortran; a wrapping could be provided via CHASM

The question now is: has someone experience with one-sided
communication, especially with SHMEM or ARMCI, and could provide some
pointers where to go next?

Our primary machine will be an SGI Altix 3700 and perhaps an Altix 4700
at the Leibniz-Rechenzentrum in Munich.

Regards!

Sebastian
.