Applying deterministic feedback suppression to reliable multicasting protocols

IP multicast is becoming the emerging infrastructure for mass delivery of information, from streaming audio and video media to multi-player games, software distributions and shared whiteboards. There have been several proposals to expand the basic unreliable service so that it can provide reliable delivery of data, in terms of ordering and loss recovery. Some of these proposals do not scale well, mainly due to feedback implosion. This is caused by excessive rate of messages arriving from receivers seeking to recover network losses. While many existing reliable-multicasting protocols seek to desynchronize receivers' feedbacks by probabilistic methods, thus circumventing the implosion problem, we show that there are certain deterministic methods, namely the Reactive Window and the Proactive Window, that guarantee implosion avoidance and provide exposure control, without incurring the overhead of excessive state and timer-based maintenance associated with probabilistic schemes. In order to demonstrate their associated performance advantages, we use both methods for building a simple reliable multicast protocol termed SDMP (Scalable Dissemination Multicast Protocol) and compare its performance with PGM (Pragmatic General Multicast), a protocol that uses probabilistic methods of de-synchronization. Our protocol SDMP: (a) takes advantage of spatial and temporal correlation of network events to deterministically control feedback implosion; (b) uses unicast feedbacks and hybrid unicast/subcast retransmissions to control delivery accuracy and exposure, thus conserving network bandwidth; (c) provides shorter arrival and recovery latencies; (d) makes use of network-based processing to detect losses and react on behalf of affected receivers; (e) accommodates local-recovery extensions; (f) has formal proof of correctness.