I'd like to apply the following to virtual/mpi. This will help the
clustering and science teams a great deal, as mpich1 contains a number of
bugs and, given the alternatives, is currently a poor choice as a default MPI
If I don't hear any objections, I'll commit this tonight.