We present MATE, a new model designed to develop communication-tolerant scientific applications. MATE employs a combination of mechanisms to reduce or hide the cost of communication from both intra-node data motion and network latency.

A key contribution of our research is the symbiotic effect of communication- hiding mechanisms, finding that their interaction yields better results than the sum of their individual contributions. We have embedded these mechanisms in a programming framework comprised of an annotation model, a source-to-source translator, and a runtime system. Co-designing these components was crucial in developing MATE as a single unified model.

We will show preliminary results of a strong scaling study of a typical 13-point stencil solver. In this study, MATE is able to reduce the majority of communication overhead on up to 128k cores (2048 nodes) of NERSC's Cori KNL platform.