数据中心网络中一种基于ECN的TCP慢启动拥塞控制策略

An ECN-based Slow-Start of TCP Congestion Control in Data Center Networks

在数据中心网络中,高带宽、低时延的链路和多对一的网络结构造成了TCP Incast吞吐率崩溃问题。现有的改进方法都关注于在TCP拥塞避免阶段改进其拥塞控制算法,而忽视了慢启动阶段中激进的指数增窗方式是造成Incast问题的关键原因。因此,该文提出了一种基于ECN的TCP慢启动拥塞控制策略(gentle slow sTart,GST),利用已有的拥塞标志位动态反馈缓存拥塞状况,以调节慢启动阶段的增窗速率。实验结果表明,该方法帮助现有的数据中心TCP协议有效地避免了并发传输中的吞吐率崩溃现象,将并发度和吞吐率分别提升了3.4倍和85倍。

TCP incast congestion happens in data center networks with high-bandwidth and low-latency,when multiple synchronized servers send data to a single receiver in parallel. The existing improved methods focuson the congestion control algorithm in the congestion avoidance phase, ignoring the aggressive exponentialincreasing window in the slow start phase that is a key reason for incast problem. Therefore, this paper proposes anECN-based slow-start of transmission control protocol （TCP） congestion control strategy, which dynamicallyfeedbacks congestion status by using the existing congestion flags to adjust the window-increasing speed during theslow start phase. Experimental results show that our approach helps the existing data center TCP effectively avoidthe throughput collapse in concurrent transmissions. The number of concurrent flow and network throughput areincreased by 3.4× and 85×, respectively.