A throughput analysis of reliable multicast transport protocols

Tree-based reliable multicast protocols are known to provide better scalability than the protocols based on pure sender- and receiver-initiated schemes. However, previous analytical work that has provided these results is based on a system model which assumes reliable control message delivery and synchronized local clocks. These assumptions are questionable simplifications, since they favor protocols using multicasted negative acknowledgments with a NAK avoidance scheme. In this paper, we extend the previous analysis by taking into account control data loss and asynchronous local clocks. We further analyze a new protocol class with particular importance, the tree-based approach with aggregated acknowledgments. In contrast to other approaches, this class provides reliability not only in case of message loss but also in case of node failures. Our results show that the additional overhead to cope with node failures is very low and therefore acceptable for reliable multicast implementations.