MaTCH : Mapping Data-Parallel Tasks on a Heterogeneous Computing Platform Using the Cross-Entropy Heuristic

We develop in this paper a new heuristic for mapping a set of heterogeneous interacting tasks of a parallel application onto a heterogeneous computing platform. The problem is well known in literature to be an NP-Hard problem. However, we propose a completely new approach based on the Cross-Entropy (CE) method. This is a new and extremely robust rare event simulation (RES) technique which may be employed to solve difficult combinatorial optimization problems (COPs). We tailor the CE method to the requirements of the problem at hand, develop a mathematical framework, and present our algorithm, MaTCH. This globally iterative randomized procedure is then compared to a previously developed genetic algorithm (GA). Through some simple experiments we prove the power of MaTCH and we get remarkable improvements in the quality of mapping. The results indicate that, when compared to the GA, MaTCH improves upon the application execution time by over a factor of 38 on a 50 node system graph. We further attest our results by performing an ANOVA test on a sample data set to prove the significance of our results.