Relaxing Heavy Congestion by State Propagation

Abstract

Interconnection network is still one of the most important key issues for building massively parallel computing systems. As a general characteristic, communication performance does not always increase as the size of network grows. Furthermore, large-scale networks suffer catastrophic performance degradation since speed of spread of congestion surpasses by far suppression speed. This paper focuses discussions on relaxation of congestion so that we can expect performance enhancement even in congested situations. This paper discusses dynamical behaviors, specifically in propagation of congestion states. When a receiver buffer becomes fully occupied, it inhibits the corresponding buffer from sending any packet to avoid loss of packet. Thus, a congested area propagates against packets' traveling direction. Based on the observation results, as the second issue of this paper, we propose a new throttling method, called State-Propagation Throttling (SPTh). The method can boost communication performance in many of typical traffic patterns in both steady and unsteady communication situations. Furthermore, this paper discusses extending the throttling method to prevent congestion from a proactive point of view. In steady communications, the proposed method improves throughput two times and latency four times. The method also improves performance of collective communication at most 1.8 times.