A model predictive control approach to flow pacing for TCP

A key challenge in the management of Internet traffic is the design of algorithms that complement well-established protocols, such as the Transmission Control Protocol (TCP), and simultaneously address their limitations. The challenge becomes greater in the context of large so-called “elephant” flows over long paths that often transition from higher to lower bandwidth connections. At these transition points either persistent queues are formed when buffers are over-provisioned or packet loss occurs when buffers are under-provisioned; both cases lead to degraded and/or highly variable end-to-end performance. Ideally, for such scenarios, the source should “learn” and set a pacing rate that matches the lower bandwidth connection. In this paper, we adopt a model-based receding horizon control strategy to design a pacing control method. Each new round-trip time (RTT) measurement is first incorporated into a linear time-varying (LTV) predictive model. Subsequently, we solve a one-step look-ahead optimization problem which finds the pacing rate which optimally trades off RTT, variance in RTT, and throughput according to the most up-to-date model. We implemented our proof-of-concept control strategy on the Linux operating system alongside the existing CoDel queuing discipline (qdisc) and HTCP congestion-control algorithm. Our preliminary results indicate significant reduction in the variances of the RTT and the throughput, resulting in more predictable performance overall.