无缓存片上网络中偏转感知的拥塞控制方法

Deflection Aware Congestion Control Mechanism for Bufferless on-Chip Networks

针对现有工作仅对高网络密集度应用的注入率进行调控,无法精确地控制网络拥塞的问题,提出一种偏转感知的拥塞控制方法——DAT.该方法使用分布式的偏转监控表来追踪网络中偏转数据包的分布,从而快速准确地确定拥塞源;并结合网络利用率确定拥塞源向网络中注入数据包的速率,以达到缓解网络拥塞提高网络整体性能的目的.在16核仿真平台上对DAT方法的评测结果显示,相比于基准无缓存No C,文中方法可以将网络偏转率平均降低8.6%,并且具有良好的可扩展性和较低的硬件开销.

Network congestion has become one of the main problems in bufferless No C, which not only degrades system throughput but also limits network scalability. Prior works only throttle the applications with high network intensity, incapable of controlling network congestion accurately. In this paper, we propose a DAT congestion control mechanism for bufferless No C, which determines the throttle sources quickly using distributed deflection monitor tables to record the deflections of network packets. And DAT takes network utilization into account when adjusting the throttling rate. Our evaluations show that DAT can reduce the deflection rate by 8.6% on average in a simulated 16 core system compared to baseline bufferless No C with good scalability and low hardware cost.