Loss Recovery at the ATM Layer for Latency-constrained Reliable Multicast

Many multicast applications have stringent time constraints, and require a data delivery guarantee. Currently, all existing reliable multicast protocols are at the transport layer. A cell discarded by a switch causes the loss of the entire packet, and eventually requires the whole packet to be retransmitted. Therefore, a small congestion problem could potentially escalate to a more serious one, and this in turn prolongs data recovery. The problem can be worse in the case of multicast, since many copies of the same packet are often retransmitted unnecessarily, reaching receivers which did not request the packet. This problem is called the exposure problem. Recovery latency would be lower if only the actual lost data were retransmitted and not the entire packet. To reduce unnecessary data retransmission and thus recovery latency of reliable multicast applications, we propose a reliable multicast protocol in which detection of data loss and retransmission are performed at the ATM layer, instead of at the transport layer. The unit of recovery is ATM cells, instead of transport-layer packets: only missing cells are retransmitted. The proposed protocol uses the ATM physical tree structure for local recovery at the ATM layer. It is thus scalable and offers low recovery latency. In addition, retransmission operates with very minimal exposure, at the cell level, that does not aggravate existing network conditions. The built-in message stability detection scheme can be used as a rate control mechanism to help alleviate congestion, and therefore reduce recovery latency. The protocol is particularly beneficial to large-scale reliable multicast applications with stringent latency requirements such as distribution of financial data, distance learning, and Internet conferencing. We have run experiments to compare the performance of ATM-layer recovery and transport-layer recovery. Simulation results show that, with ATM-layer recovery, the average retransmission delay is significantly lower, and the connection throughput is higher under congestion.