o Message Passing != Distributed Computing.
This is at the heart of the question "Should I Use PVM or MPI?" If you are doing distributed computing then PVM is the natural choice. If only message passing, then MPI is the natural choice. We should go beyond "theological" discussions and work on the best way to merge the two.
o The library approach is too confining!
Distributed computing libraries have served us well to date. But, we need some help from compiler writers. We should be able to say send (structure foo, task B, tag alpha) and not worry about types. Basically, I find derived datatypes in MPI, for example, too confusing or unwieldy to use effectively.
o We need a Distributed Computing Engineering Task Force (DCETF) that is similar to the IETF (Internet Engineering Task Force).
We (the message passing/distributed computing communinity) need to work
together to "engineer" the next interfaces. MPI-1 worked well
because implementations were being undertaken as the specification was
being formalized.
MPI-2, on the other hand, defined a standard interface without
any existing implementations. This is standard by fiat, which is
destined to fail because the details could not be seen up front.
Complete MPI-2 implementations are years away. Requirements
*will* change in that period of time.
A DCETF can provide a flexible environment where ideas and implementation
can come together to build a standard distributed computing environment.
MPI and PVM are evolving.
A DCETF provides a framework for MPI and PVM to evolve together so
that end-users can have long-term stable environments. PVM and MPI
complement each other, we need to build on this, take only
the best of PVM and MPI, and go forward.