End-to-end TCP-compatible Backpressure Routing

Backpressure constitutes a well-known throughput-optimal packet scheduler. Backpressure scheduling decisions are taken based on commodity queue build-ups at the network nodes, and constant status updates among neighbors. Nonetheless, it is also well known for its incompatibility with TCP, which uses congestion control to avoid queue build-ups, while also being sensitive to packet timeouts and their delivery order. These factors limit the use of backpressure in the local packet forwarding plane, e.g., within switches. The present work presents a breakthrough in combining TCP and backpressure at the network layer, overcoming any compatibility issues. The key-idea is to apply the Lyapunov framework cumulatively, over a large set of router hardware clock ticks. This approach allows for the throughput-optimal backpressure decisions to be expressed via regular traffic engineering routing rules, instead of explicit per-packet forwarding directives. Moreover, these throughput-optimizing routing rules are derived using common metrics logged at commercial routers, and not via queue build-ups. Simulation results validate the analytical findings and demonstrate significant throughput gains over state-of-the-art related approaches.