Fine-grained congestion control for multipath TCP in data center networks

Abstract

In data center networks, MultiPath TCP (MPTCP) obtains both higher network utilization and fairer allocation of capacity by exploring multiple paths simultaneously. However, MPTCP experiences more queue oscillation in switch buffer with the increasing of the number of subflows. When using explicit congestion notification as the congestion indication to track the queue length in switch buffer, it is difficult for MPTCP to capture the accurate congestion state, resulting in wrong behavior in congestion control. Therefore, we propose an enhanced MPTCP protocol, namely, advanced MPTCP (AMP), which adjusts the time granularity of the congestion detection and control under a different number of subflows. The test results show that compared with Linked Increases Algorithm and eXplicit MultiPath, AMP achieves lower latency for small flows and higher throughput for large flows.