Traffic Pattern Adaptive Hybrid Electrical and Optical Switching Network for HPC System

Abstract

As the performance requirements and scale of high-performance computing (HPC) systems grow rapidly, systems with such a large number of computing nodes present significant challenges in network architecture design because their peak performance is increasingly limited by communication rather than computation. Moreover, certain HPC application always shows some specific communication patterns, and different topologies have various performances with these patterns. To implement a traffic pattern adaptive HPC system, we propose a topology reconfigurable hybrid network architecture based on optical circuit switching (OCS), which could be adjusted into different classical topologies. And an algorithm of topology calculating and a permutation matrix-based optical switch configuration mechanism are proposed to construct network. We discuss network performances with limited OCS switches occupation, and results show that traffic pattern adaptive network improves the performance of latency and throughput to a large extent.