Accelerating MPI Message Matching by a Data Clustering Strategy

Message Passing Interface (MPI) is one of the most popular parallel programming models for high-performance computing. In MPI, message matching operations are in the critical path of communication, which could adversely affect the application performance. MPI libraries typically use a linked list data structure to maintain early posted receives and unexpected messages for matching operations. However, they perform poorly at scale due to long message queue traversals. In this paper, we propose an MPI message matching mechanism based on K-means clustering that considers the behavior of the applications to categorize the communicating peers into clusters, and assign a dedicated queue to each cluster. The clustering is done based on the number of queue elements each communicating process adds to the posted receive or unexpected message queue at runtime. The proposed approach provides an opportunity to parallelize the search operation for different processes based on the application‘s message queue characteristic. The experimental study with real applications confirm that the proposed message matching approach reduces the number of queue traversals, the queue search time, and the application runtime.