This paper describes an efficient implementation of MPI on the Memory-Based
Communication Facilities; Memory-Based FIFO is used for buffering
by the library, and Remote Write for communication with no
buffering.
The Memory-Based Communication Facilities are software-based communication
mechanisms, with off-the-shelf Ethernet hardware.
They provide low-cost and highly-functional primitives for remote
memory accesses.

The performance of the library was evaluated on a cluster of workstations
connected with a 100Base-TX network.
The round-trip time was 71 us for 0 byte message, and the
peak bandwidth was 11.86 Mbyte/s in full-duplex mode.
These values show that it is more efficient to realize the message
passing libraries with the shared memory model than with the message
passing model.