MPCC

Abstract

Multipath transport, as embodied in MPTCP, is deployed to improve throughput and reliability in mobile and residential access networks, with additional use-cases including spreading load in data centers and WANs. However, MPTCP is fundamentally tied to TCP Reno's legacy AIMD algorithm, and significantly lags behind the performance of modern single-path designs. Consequently, MPTCP fails to achieve high performance in many real-world environments. We present MPCC, a high-performance multipath congestion control architecture. To achieve our combined goals of fairness and high performance in challenging environments, MPCC employs online convex optimization (a.k.a. online learning). In experiments with a kernel implementation on emulated and live networks, MPCC significantly outperforms MPTCP.

Machine Learning vs. Classical Methods: Evaluating Congestion Control Performance in the Transport Layer

“…Congestion control in a multipath setting would involve applying regular TCP congestion control mechanisms to each subflow. However, this approach is not as efficient as MPTCP ' [9], which utilizes online learning and has been implemented in the Linux kernel. This method has been tested under various network conditions and topologies.…”

A Survey on Congestion Control and Scheduling for Multipath TCP: Machine Learning vs Classical Approaches

“…Traditional AIMD used in TCP shows poor performance adaptation in terms of network state-changing situations in MPTCP. Gilad et al [44], have presented a method named MPCC that uses online learning. Their implementation has been performed in Linux kernel and the method has been tested on different network conditions and many different network topologies.…”

PRISM based Transport

“…Congestion control protocols, such as PCC, obtain fairness by using online learning [12], [13], [18], [33], [35], [36]. They constantly get feedback from the network, in the form of loss ratio and RTT, regarding the current congestion and adjust the sending rate of the endpoints.…”

“…The LMMF vector doesn't consist of just the max min rate (arg max x 0 ) but also the maximums of any other rate (x i ) given the values of the lower rates (x 0 , ..., x i−1 ). Therefore, LMMF achieves Pareto efficiency [28], [30], [33] and always exists (as long as the number of flows is finite) [34].…”

“…The loss coefficient β ≥ 0 sets a threshold on random loss tolerance. The latency coefficient γ > 0, corresponds to a theoretical maximum number of PCC guarantees equal sending rate for competing flows (fair-share) and was extended to multi-path connections, by MPCC [33], which was proved to converge to LMMF.…”