Directional Gossip: Gossip in a Wide Area Network

A reliable multicast protocol ensures that all of the intended recipients of a message m that do not fail eventually deliver m. For example, consider the reliable multicast protocol of [10], and consider a message m, sent by process p1, that is intended to be delivered by p1, p2, and p3. We impose a directed spanning tree on these processes that is rooted at the message source. For example, for m we could have the directed spanning tree p1 ! p2 ! p3. The message m propagates down this spanning tree and acknowledgments of the receipt of m propagate back up the tree. A leaf process in this tree delivers m when it receives m, and a non-leaf process delivers m when it gets the acknowledgment for m from all of its children. If a non-leaf process (say, p1) does not get an acknowledgment for m from one of its children (here, p2), then it removes the child from the tree and \adopts" that child's children (here, p3). The process sends m to the newly-adopted children and continues the broadcast. A similar monitoring and adoption approach is used to recover from the failure of the root of the tree. Reliable multicast protocols are intended for local area networks. Unfortunately, most implementations of reliable multicast do not scale well to large numbers of processes even when all are in the same local area network [3]. For example, with the protocol given above, the sender cannot deliver its own message m until it knows that all non-failed processes have already delivered m. The latency can be reduced by using a bushy directed spanning tree, but doing so increases the overhead of some processes, where by overhead we mean the number of messages a process sends and receives in the reliable multicast of a single m. As the number of processes increases, either the latency or the overhead at some processes increases. Hence, when a multicast is to be sent to a large number of processes or processes located on a wide area network, a protocol like IP Multicast [4] that has been speci cally designed for these cases is preferable even though it is not as reliable as reliable multicast. More recently, gossip-based protocols have been developed to address scalability while still providing high reliability of message delivery. These protocols, which were rst developed for replicated database consistency management in the Xerox Corporate Internet [5], have been built to implement not only reliable multicast [3, 7] but also failure detection [11] and garbage collection [12]. Gossip protocols are scalable because they don't require as much synchronization as traditional reliable multicast protocols. A generic gossip protocol running at process p has a structure something