Two Approaches for High Concurrency in Multicast-Based Object Replication

This report presents a replica control protocol for atomic objects. The protocol is derived from an atomic broadcast primitive, and places constraints on the delivery of messages to provide a consistent message order among sites. Several heuristic techniques are proposed to reduce the latency of message delivery, for two types of orders. Messages are delivered either in the same order for all sites, or in an order semantically equivalent to this unique ordering. The equivalence relation is based on the commutativity property of operations on objects, i.e: two deposit operations commute. The protocol uses a reliable causal multicast primitive, and is fully distributed. The rst set of heuristics is based on a voting scheme, and delivers messages in a unique order. Totally ordered atomic multicast can be built on top of a reliable causal multicast by waiting until each processor in the group has multicast a message, inserting them in a causal graph, and then delivering the roots of this graph. This latency can be reduced with a voting scheme that allows messages to be inserted in a total order without waiting for a message from all group members. Dolev, Kramer and Malki proposed such a protocol, but it is more restrictive than it needs to be, and can be extended for a wider range of operating conditions. In particular, we show how to deliver a message even if it is acknowledged by less than half of the processors. If there remains some undelivered messages (not yet placed in the total order), a second set of heuristics are available to a processor. If messages correspond to operations on objects, it can use local and type speciic information to minimize delivery constraints. More speciically, if operations commute, it is possible to initiate some of them early, provided that the execution is equivalent to a linear sequence of operation invocation, the same for all sites. The commutativity of operations depends on the state of the data, thus allowing more concurrency than otherwise possible. We present a general algorithm for ordering operations provided that messages are causally ordered, and give two heuristics to make it practical. Finally, we present a performance evaluation of the protocols based on discrete-event simulation. Our multicast primitives perform well, and adapt better than previous work to various network conngurations. It is also more scalable.