PVM Implementations of Fx and Archimedes

We have ported Fx, an experimental HPF compiler, and Archimedes, a
parallelizing finite element methods compiler to PVM.  These ports
have allowed us to run Fx and Archimedes programs in several different
workstation cluster environments, including those with switched-FDDI
and HIPPI networks, as well as on the Intel Paragon, and the Cray T3D.
This talk will explain how the PVM ports work and comment on the
portability of PVM programs, and on PVM as a compiler target and user
interface.  

Generally, we are reasonably happy with the portability of PVM,
although every non-workstation port has required some modifications.
Debugging and performance testing of PVM programs on workstation
environments is far easier than on commercial parallel supercomputers.
However, it appears that the price to pay for portability and ease of
use is performance - PVM's performance on each of our targets is
markedly less than with the native message passing system or with
systems specialized for our compilers.  Finally, we have found it
difficult to hide the details of PVM from our users.