Performance Study of Non-blocking Collective Communication Implementations Toward Adaptive Selection

Abstract

Non-blocking collective communication is a key technique for enabling overlap of collective communication and computation to achieve higher scalability on large scale parallel computers. There can be many types of implementation methods for non-blocking collective communications. To sufficiently obtain effect of overlap, appropriate method should be chosen. Therefore, it is important to study characteristics of each method. In this paper, we explain a characteristic of each method of implementation and and evaluate the characteristic by performing the experiment using a simple test program. In addition to that, this paper proposes an implementation with lower overhead, and evaluates the performance.