Compiling Communication-Efficient Programs for Massively Parallel Machines

摘要

This paper presents a method of generating parallel target code with explicitcommunication for massively parallel, distributed-memory machines. The source programs are shared-memory parallel programs with explicit control structures. The method contains four main parts: (1) analysis: extracting syntactic reference patterns from a program with shared address space, (2) pattern matching: selecting appropriate communication routines, (3) scheduling: placing these routines in appropriate locations in the target program text, and (4) synchronization: setting up correct conditions for invoking these routines. We use an explicit communication metric to guide the selection of data layout strategies.