I'm away from a computer right now so I don't have the specifics, but we saw some testbake results from UH today of mpich2 under slurm that were not run properly - it ran 16 copies of skampi instead of 1 16-node job, so the output was very skewed (and completely mis-parsed).
Can you check your mpich2 compile / link settings to ensure that you're linking against the slurm pmi library properly? -jms
