Scheduling on parallel processing systems using parallel primitives

An approach for integrating task and parallel architecture characteristics is presented, with the objectives of easing the burden of parallel programming on the user and of increasing the system efficiency through more informed embedding, partitioning, and scheduling. The proposed approach uses the concepts of parallel primitives and a primitives table that serves as the knowledge base for the system. Parallel primitives are parallel SIMD computations and can be used to write parallel programs. The primitives table stores the specifications of each primitive. In addition, for each primitive several alternative implementations corresponding to alternative network structures are provided and stored in the table. A layered implementation of a parallel processing system with parallel primitives is presented. The authors also examine the positive implications of the primitives table on scheduling, as well as on determining the optimal machine configurations for tasks using precedence graphs. The latter problem is shown to be NP-complete. However, a linear algorithm is found for the case in which the precedence graph is a tree.<>