A Machine Learning Based MPTCP Subflow Control Algorithm for Multi-access Heterogeneous Networks

Ever growing number of devices and the demand for applications are generating tremendous network traffic today. In the edge network environment, due to the network instability caused by device movement, the single network cannot meet its throughput requirements. Hence, using multi-access network to transmit data packages has become a way to improve throughput. Multi-path TCP utilizes multiple channels to transmit data packages simultaneously. However, MPTCP is unable to control subflows adaptively, and the disorder of data packages caused by difference between paths leads to its poor performance compared with the single path transmission. Therefore, a new MPTCP subflow control algorithm (MSCA) based on Support Vector Machine (SVM) prediction model is proposed. The SDN controller is utilized to continuously monitor the network status and predicts the impact factor of the network based on the path parameters monitored. Then, according to the impact factor, the subflow configurations is dynamically adjusted by the system for each user to improve the average throughput. The system is implemented on the network emulator platform Mininet-WiFi. Emulation results shows the proposed MSCA improves the average throughput compared with traditional scheme. MSCA also responses quickly to the network changes.