GFlow: GNN-Based Optimal Flow Scheduling for Multipath Transmission With Link Overlapping

Multipath TCP (MPTCP) is considered as a solution capable of addressing the growing demand for bandwidth. However, the existing MPTCP mechanisms make flow scheduling based on coarse-grained end-to-end network states, which prevents MPTCP from better aggregating the bandwidth of multiple paths. Besides, link overlapping may occur between different MPTCP connections, which results in multiple subflows competing for bandwidth of the shared link. In this paper, we propose GFlow, a Graph Neural Network (GNN) based Deep Reinforcement Learning (DRL) algorithm, to make optimal flow scheduling for multipath transmission with link overlapping. Specifically, we formulate the flow scheduling problem as a problem of maximizing overall throughput by taking both bottleneck bandwidth and shared bandwidth into consideration. To support accurate network state perception, GFlow utilizes In-band Network Telemetry (INT) to collect real-time and fine-grained network states. Taking these states as input, the DRL agent with GNN integrated fully learns the relationships among links, paths (subflows), and MPTCP connections. In this way, GFlow is able to make optimal flow scheduling decisions according to the network states. We build a P4-based multipath transmission system and carry out extensive experiments to evaluate the performance of GFlow. The results show that GFlow outperforms the baseline multipath transmission mechanism in both homogeneous scenario and heterogeneous scenario, improving the average overallthroughput while reducing the average round trip time (RTT).