Improving multipath TCP for latency sensitive flows in the cloud

Recent cloud datacenter network usually has multiple transmission paths for any pair of servers to increase the aggregate bandwidth and to improve the robustness of failure recovery. Therefore, Multipath TCP (MPTCP) is becoming one of the most important transport protocols exploring the multipath resource to improve the performance of cloud network. It is obvious that MPTCP could significantly increase the bandwidth for the throughput-oriented long flows by using multiple paths concurrently. However, for many latency-sensitive applications that require low latency other than high bandwidth, MPTCP performs poor. In this paper, we propose a set of extensions for MPTCP (called MPTCP-L) to improve its performance on the latency. Particularly, MPTCP-L replicates packets of latency-sensitive flows on two MPTCP subflows to proactively reduce retransmission delay of lost packets. Besides, we propose a rapid packet retransmission scheme that takes only 1-2 RTTs to recovery the lost packets for the TCP subflows of MPTCP. We conduct large scale simulations and testbed experiments based on real traffic load to evaluate MPTCP-L, and the results prove that MPTCP-L outperforms the standard MPTCP for latency-sensitive flows, and the induced redundancy overhead is negligible.6 page(s