RBPCCM: Relax Blocking Parallel Collective Communication Mechanism Base on Hardware with Scalability

With the development of parallel computation, the scale of high performance computing system increases dramatically and the collective communication has become its bottleneck. The collective communication with the hardware support has the relatively high performance. However, scalability of collective communication is always a crucial problem, because the number of nodes involved is not fixed. This paper proposes the Relax Blocking Parallel Collective Communication Mechanism (RBPCCM) to improve the performance of the collective communication in parallel computation. This mechanism, cooperating hardware and software, implements the scalable collective communication by distributing collective resource allocation numbers. Furthermore, RBPCCM supports the implementation in various scales of endpoint, unconstrained by the interconnect network topology. A functional simulation model is built based on the system of Sunway Taihu Light to verify the correctness and scalability of this proposed method. The implementation of RBPCCM prototype is built based on the network interface, and a FPGA platform is constructed for performance test. It is testified that RBPCCM has the improvement as regards to delay performance from 2.4 to 37 times, compared with the Point-to-Point communication based on software.