MPTD: optimizing multi-path transport with dynamic target delay in datacenters

Abstract

In contemporary datacenter networks, various applications experience the challenge of incast occurrences. Data-intensive applications, particularly in distributed scenarios such as distributed deep learning, frequently encounter incast. While single-path transport protocols are widely deployed in datacenter networks, the available bandwidth is often severely underutilized. MP-RDMA, as a state-of-the-art multi-path transport protocol, has been implemented; however, it faces performance issues during large-scale incast communications. MP-RDMA employs insufficient congestion feedback and fixed rate adjustment for congestion management, making the scheme hardly scalable. In this paper, we introduce MPTD, an RTT-based multi-path transport protocol. MPTD encompasses two primary aspects: (1) a novel congestion control algorithm based on the setting of target delay; (2) an optimized packet scheduling mechanism. Specifically, the receiver dynamically updates the target delay in real-time based on network conditions, and the sender adjusts the sending rate accordingly. A categorized implementation for paths is added to achieve a more accurate identification of path characteristics. Our evaluation demonstrates that, in comparison to MP-RDMA, MPTD can achieve speedups of 1.35x/1.40x in flow completion time for two traffic patterns.