PPS: A Low-Latency and Low-Complexity Switching Architecture Based on Packet Prefetch and Arbitration Prediction

摘要

Interconnect networks increasingly bottleneck the performance of datacenters and HPC due to ever-increasing communication overhead. High-radix switches are widely deployed in interconnection networks to achieve higher throughput and lower latency. However, network latency could be greatly deteriorated due to traffic burst and micro-burst features. In this paper, we propose a Prefetch and prediction based Switch (PPS) which can effectively reduce the packet delay and eliminate the effect of traffic burst. By using dynamic allocation multiple queueing (DAMQ) buffer with data prefetch, PPS implements concurrent write and read with zero-delay, thus implementing full pipeline of the packet scheduling. We further propose a simple but efficient arbitration scheme, which completes a packet arbitration within one clock cycle meanwhile maintaining higher throughput. Moreover, by predicting the arbitration results and filtering the potential failed requests in the next round, our scheduling algorithm demonstrates indistinguishable performance from the iSLIP, but with nearly half of the iSLIP's area and 36.37% less logic units (LUTs). Attributing to the optimal schemes of DAMQ with control data prefetch and two-level scheduling with arbitration prediction, PPS achieves low-latency and high throughput. Also, PPS can easily extend the switching logic to a higher radix for the hardware complexity grows linearly with the number of ports.