Automatic Discovery of Implementation Rules for Fast GPU + MPI Operations

Developing a high-performance implementation of a distributed computational kernel for high-performacing computing is increasingly challenging. Systems are composed of heterogenous computational resources, and limited communication performance demands an asynchronous application design. Even if high-performance computation and communication libraries are available. the challenge becomes the best coordination of the provided operations to create an optimal result. This work presents a system that automatically generates design rules for a high-performance implementation of a compound operation provided as a dependence graph. The system searches among valid schedules to determine the fastest arrangement of operations. A post-processing step on the results of the search yields interpretable design rules. The fast implementation can be used directly, or experts can use the design rules to create a high-performance implementation.