Optimal deterministic timeouts for reliable scalable multicast

Reliable multicast suffers from the problem of feedback implosion. To achieve scalability, the number of receivers sending feedback in case of loss must remain small. However, losses experienced by different receivers are strongly correlated, since they share resources in the multicast tree. We present DTRM (deterministic timeouts for reliable multicast), a distributed algorithm to compute optimal deterministic timeouts for each receiver in a multicast tree as a function of the tree topology and sender-to-receiver delays. DTRM has several desirable properties. First, the computation of the timeouts is entirely distributed; receivers and intermediate nodes only rely on local topology information. Second, NACK implosion is provably avoided for a single loss anywhere in the tree if delay jitter is bounded. Third, feedback information does not need to be processed by intermediate nodes, and receivers do not have to collaborate. We foresee two possible uses for DTRM. In networks providing hard delay bounds, timeouts can be computed once at session set-up time. In networks with unbounded delays, such as the Internet, timeouts can be adaptively recomputed in response to changes in estimated round-trip times.