A Unified Compiler Framework for Work and Data Placement

In parallel programming, the nature of the distribution of the data over the processors, and the assignment of work to the processors in the system, strongly influence the performance of the program. In a simple parallel programming environment this information is entirely specified by the user, but this places a heavy burden on the user. Development of parallel programs is much easier and robust if the user can leave part of the placement to the compiler. To support this, the method of specifying distributions must allow partial placement specification. The compiler must then find the optimal choices for the unspecified placements. In this paper we present a compiler framework that takes a program with partial work and data placement information, and transforms it into an explicit parallel program. The unspecified placements are chosen to minimize the number of required communication steps.