Exploiting Client Inference in Multipath TCP Over Multiple Cellular Networks

In this article, we present a client-based multipath TCP (cMPTCP) framework that can cope with path heterogeneity across multiple LTE networks deployed by different operators. cMPTCP is the first work to apply the mobile client's inference based on our observation that an out-of-order queue in an MPTCP receiver (i.e., mobile client) builds up with increasing heterogeneity between different subflows and can significantly degrade throughput performance compared to that achieved by single-path TCP via the best path. For agile and accurate control for the existing MPTCP architecture, we argue that mobile clients are the best network entities to infer the bottleneck state of an end-to-end MPTCP connection. Our proposed cMPTCP consists of two main components: client-initiated multipath selection and subflow-level rate control mechanisms. These mechanisms are compatible with the default MPTCP architecture implemented in the Linux kernel. For performance evaluation, we prototype a complete LTE network testbed consisting of mobile phones, LTE base stations, LTE cores, and application servers. Our experimental results confirm that cMPTCP outperforms existing server-side schedulers in practical dynamic channel conditions.