Analysis and improvement of the latency-based congestion control algorithm DX

摘要

In data center networks, many congestion control algorithms have been proposed to achieve low latency, in which DX is a famous latency-based congestion control algorithm. Although the advantages of DX have been confirmed by experimental results, the theoretical analysis of DX is absent. Accordingly, some drawbacks of DX under special environments are unexplored. In this paper, we conduct fluid-flow analysis over DX, deduce sufficient conditions for the stability of DX and reveal its special stable state when with a large number of concurrent flows or the small Round Trip Time (RTT). Analytical results uncover two problems of DX: 1) it has poor throughput when either the base RTT is very large or the number of flows is relatively small; 2) it suffers from large queueing delay when either the base RTT is relatively small or the number of concurrent flows is large. The reason is that the congestion window is always required to be reduced even if it already is the smallest value 1, owing to the small RTT in data centers. In order to solve the problem that DX cannot continue to reduce the sending rate in the special stable state, we propose DX+. DX+ regulates the time interval of sending packets when the congestion window is 1 but the link is still congested. Extensive simulation results verify our analytical results and show that DX+ can further reduce queueing delay without losing throughput under special stable conditions compared to DX.