Entropy Throttling: Towards Global Congestion Control of Interconnection Networks
Takashi Yokota, Kanemitsu Ootsu, Takeshi Ohkawa · 2015
Importance of interconnection networks is continuously increasing as the number of processing elements in massively parallel computers grows. Wide spectrum of efforts in research and development for effective and practical interconnection network methods are reported, however, the problem is still open. This paper focuses discussions on congestion control, which intends to minimize congestion in order to maximize performance (maximal throughput and minimal latency). The major contribution of this papers is to clarify the effectiveness of Entropy Throttling (EntTh) and to propose an improved method. This paper first introduces Entropy Throttling whose foundation is the packet entropy. The packet entropy is proposed to represent the degree of congestion situation. It can properly represent phase transition phenomena between congested and uncongested situations. This paper then proposes an enhanced method of EntTh by introducing a hysteresis feature for further improvement. Comprehensive performance results of the originated and enhanced EntTh methods are unveiled, where steady/unsteady communications under various traffic patters are assumed. The enhanced version of EntTh accelerates collective communication performance at most 1.5 times faster than non-throttled cases.