This could be clarified by stating that implementations may auto-start
the daemon(s) in the singleton case, or may require that the daemon(s)
to be started independently. I think this is acceptable. The issue
we're hitting on here is: what is the MPI environment? Is the LAM/PVM
daemon part of it or part of MPI? I think they're not part of MPI, they
both do more than just run MPI jobs. Just as the MPI-singleton wouldn't
be required to boot UNIX, I think it's acceptable not to require it to
boot other non-MPI environments it may depend on. With this, we can
then leave it up to the implementations to decide which part they
consider "MPI environment" (and thus auto-start with the singleton) and
which parts they exclude. This doesn't just apply to LAM/PVM, it also
applies to resource managers. If it helps, just consider LAM/PVM daemons
to be "resource manager"-like.
Raja Daoud Hewlett-Packard Co.
email@example.com Convex Division