Predictive Queue-based Low Latency Congestion Detection in Data Center Networks.

End-to-end congestion control in lossless data center networks (DCN) depends on congestion detection. However, many current congestion control mechanisms ignore the mutually reinforcing relationship with congestion detection and flow identification, resulting in wasted link resources. For accurate congestion detection, the currently popular ternary congestion detection (TCD) technique sets three separate port states. However, the flow needs to go through multiple hops before deciding the flow category, which leads to serious delays. In order to achieve accurate congestion detection and low delay flow identification, we propose a predictive queue-based low latency congestion detection, called hop-by-hop flow identification (HFI) based on congestion detection. HFI predicts the dynamic queueing threshold based on the data transmission rate to provide timely and proactive congestion detection, then combined it with the congestion information notification from the next hop to identify the congested flow hop by hop. Experimental results show that DCQCN+HFI obtains shorter flow completion time, and improves link utilization by 55%, 7.5% compared to DCQCN, DCQCN+TCD.