An architectural style for high-performance asymmetrical parallel computations

Researchers with deep knowledge of scientific domains are becoming more interested in developing highly-adaptive and irregular (asymmetrical) parallel computations, leading to development challenges for both delivery of data for computation and mapping of processes to physical resources. Using software engineering principles, we have developed a new communications protocol and architectural style for asymmetrical parallel computations called ADaPT.Utilizing the support of architecturally-aware middleware, we show that ADaPT provides a more efficient solution in terms of message passing and load balancing than asymmetrical parallel computations using collective calls in the Message-Passing Interface (MPI) or more advanced frameworks implementing explicit load-balancing policies. Additionally, developers using ADaPT gain significant windfall from good practices in software engineering, including implementation-level support of architectural artifacts and separation of computational loci from communication protocols.