Yiannis Thomas successfully presented his Ph.D. thesis, with title "Multipath Internet Transport", on Wednesday July 18th
The proliferation of smartphones, with their multiple interfaces, and data servers, with their high-performance interconnection networks, has revived interest in multipath transport protocols. Multipath-TCP (MPTCP), the multipath extension of TCP, is currently available in the Apple iOS and Linux operating systems, enabling bandwidth aggregation, load balancing, and resilience to failures and disconnections due to mobility. However, the deployment of multipath transport is challenged by the address-based TCP/IP communication, which does not facilitate the seamless establishment of multiple paths among two end-points, and by the distributed hop-by-hop TCP/IP routing, which does not ensure the disjointness of the paths. Even when multiple paths are deployed, the use of many subflows is both a blessing and a curse for multipath TCP/IP protocols, as they tend to grasp an unfair share of bandwidth, thus becoming unfriendly to single-path TCP flows. The latest congestion control algorithms for MPTCP attempt to equalize the cumulative subflow throughput with the throughput of the fastest single-path flow in the same link, thus exchanging performance for TCP-friendliness. While TCP-friendly in the long run, these approaches exhibit high throughput convergence latency, thus being effective only for long-lived flows.
Our contribution to multipath transport is twofold. First, we introduce the multipath multisource Transport Protocol (mmTP), a transport-layer protocol that offers reliable multipath and multisource content delivery in the Publish-Subscribe Internet (PSI) architecture. mmTP increases the utilization of network processing resources, exploits on-path and off-path caching and does not require additional state at routers, or complex signaling during connection establishment. Second, we propose a novel hybrid multipath congestion control algorithm that enhances resource utilization through greedy friendliness, a design that meets the TCP-friendliness constraint only when is needed. The hybrid congestion control scheme consists of the novel end-to-end Normalized Multiflow Congestion Control (NMCC) algorithm, which offers instant convergence to TCP-Friendliness, and an in-network topology management module, that provides disjoint paths when possible and notifies end-users about shared bottlenecks otherwise. We finally discuss the integration of the proposed designs with the TCP/IP architecture: mmTP through Software Defined Networking (SDN) and NMCC through MPTCP.