Content area
The issue of synchronization is of particular importance in the domain of parallel and distributed computation. Synchronization between processes establishes the dependency between processes. This dependency relation plays an important role in the level of parallelism and speedup that an application can achieve. In this dissertation, the nature of synchronization primitives is identified and the types of synchronization primitives that are useful in such environments are discussed. Furthermore, we go beyond the syntax of a program and investigate the notion of semantics: the meaning of operations and programs. This notion is demonstrated to lead to a better performance in parallel and distributed programs. First, two kinds of semantics, operation semantics and program semantics, are identified and the model for computation is defined. Operation semantics is then applied to improve the time warp mechanism for virtual time computations, which are extremely useful in parallel discrete event simulation. The same approach is also employed in distributed rollback recovery protocols with optimistic message logging schemes, which are used to provide fault-tolerance for long-lived parallel programs. In both applications, operation semantics is used to reduce the dependency between participating processes as a result of synchronization, and both are based on message passing systems. For the development of parallel and distributed applications on shared memory systems, the model of distributed shared memory system is advocated and described. Object-oriented distributed shared memory systems as well as non-atomic memory systems are shown to be useful. Useful synchronization primitives for non-atomic memories are studied and then implementation issues are addressed. Program semantics is exploited for different applications executing under these systems. Finally, an evaluation platform, Maya, is developed and used to demonstrate the usefulness of non-atomic and object-oriented distributed shared memory systems.