Reducing transport latency for short flows with multipath TCP

Abstract

Multipath TCP (MPTCP) has been an emerging transport protocol that provides network resilience to failures and improves throughput by splitting a data stream into multiple subflows across all the available multiple paths. While MPTCP is generally beneficial for throughput-sensitive large flows with large number of subflows, it may be harmful for latency-sensitive small flows. MPTCP assigns each subflow a congestion window, making short flows susceptible to timeout when a flow only contains a few packets. This condition becomes even worse when the paths have heterogeneous characteristics as packet reordering occurs and the slow paths can be used with MPTCP, causing the increased end-to-end delay and the lower application Goodput. Thus, it is important to choose the appropriate subflows for each MPTCP connection to achieve the good performance. However, the subflows in MPTCP are determined before a connection is established, and they usually remain unchanged during the lifetime of that connection. To address this issue, we propose DMPTCP, which dynamically adjusts the subflows according to application workloads. Specifically, DMPTCP first utilizes the idea of TCP modeling to estimate the latency on the path under scheduling and the data amount sent on the other paths simultaneously, and then decides the set of subflows to be used for certain application periodically with the goal of reducing completion time for short flows and achieving a higher throughput for long flows. We implement DMPTCP in a Linux server and conduct extensive experiments both in NS3 and in Linux testbed to validate its effectiveness. Our evaluation shows that DMPTCP decreases the completion time by over 46.55% compared to conventional MPTCP for short flows while increases the Goodput up to 21.3% for long-lived flows.