Semantically reliable group communication

Current usage of computers and data communication networks for a variety of daily tasks, calls for widespread deployment of fault tolerance techniques with inexpensive off-the-shelf hardware and software. Group communication is in this context a particularly appealing technology, as it provides to the application programmer reliability guarantees that highly simplify many fault tolerance techniques. It has however been reported that the performance of group communication toolkits in large and heterogeneous systems is frequently disappointing. Although this can be overcome by relaxing reliability guarantees, the resulting protocol is often much less useful than group communication, in particular, for strong consistent replication. The challenge is thus to relax reliability and still provide a convenient set of guarantees for fault tolerant programming. This thesis addresses models and mechanisms that by selectively relaxing reliability guarantees, offer both the convenience of group communication for fault tolerant programming and high performance. The key to our proposal is to use knowledge about the semantics of messages exchanged to determine which messages need to be reliably delivered, hence semantic reliability. In many applications, some messages implicitly convey or overwrite other messages sent recently before, making them obsolete while still in transit. By omitting only the delivery of obsolete messages, performance can be improved without impact on the correctness of the application. Specifications and algorithms for a complete semantically reliable group communication protocol suite are introduced, encompassing ordered and view synchronous multicast. The protocols are then evaluated with analytical and simulation models and with a prototype implementation. The discussion of a concrete application illustrates the resulting programming interface and performance.