Reliable streaming protocol for lossy networks

This paper introduces Rest, a reliable streaming protocol for lossy networks. Rest ensures full reliability while recovering losses as soon as possible thanks to the proactive injection of redundancy packets encoded following an on-the-fly scheme. It dynamically adapts the sending of codes depending on the estimation of the packet error rate with periodic acknowledgments to limit feedback dependency and protocol overhead. Results show that data are smoothly delivered to the receiving application with minimum overhead when errors are uniform. For systems with limited processing capacity, we propose to use a bounded encoding window to deliver data more uniformly while limiting decoding matrices size. We study the performance of Rest under different network conditions and highlight the underlying trade-offs behind each system parameter. We show that an optimal acknowledgement frequency can be estimated to minimize overhead while meeting system requirements in terms of delivery delay and computational power.