Published in Computer Communications, 2012
Today’s dominant Internet transport protocol – TCP – is known to achieve a low bandwidth utilization on lossy large bandwidth-delay product paths. As such paths become common in the Internet, TCP’s inefficiency will be the main bottleneck in deploying high-rate, low-latency communications services such as tele-presence, free-viewpoint video or virtual reality. In this paper, we present PRRT – Predictably Reliable Real-time Transport, a protocol layer that efficiently supports these services. PRRT is based on a hybrid error correction that combines proactive and reactive reliability and dynamically adjusts the amount of redundancy to path properties. By relying on an analytical model of the path, it achieves a near-optimal packet-erasure correction under service-specific end-to-end delay. As a result, it simultaneously achieves a low latency and a high bandwidth utilization. We describe key features of the PRRT: (1) adaptive block-erasure coding and (2) predictable reliability based on a statistical performance model. We implement PRRT in the Linux kernel and evaluate its performance over a range of large bandwidth-delay product paths under a strict service-specific delay.