Yet another title possibility:

Unbundled communication in compiler-generated parallel programs.

Abstract: Supercomputer message passing models typically bundle a rich
set of services into their send and receive statements.  These
services include the basic data transfer service, as well as
message-level synchronization, flow control, selective receives, and
buffer management. While this rich functionality greatly simplifies
the job of the message passing programmer, it adds to communication
overheads.  Further, if the message passing program is generated by a
parallelizing compiler, many of these services are not needed.  We
propose an approach for building communication systems based on the
idea of unbundled communication.  We identify the basic services
provided by message passing systems, show how to combine these basic
services to form simpler and more efficient message passing models
that can be exploited by parallelizing compilers, and present measured
performance results on the Cray T3D to validate the approach.