Performance optimization for distributed-shared-data systems

Abstract

This thesis introduces an all-user-level-library approach to using distributed shared data in a network-of-workstations environment, which has a relatively high-latency, low-bandwidth communication network. The approach is based upon managing arbitrary distributed user-defined dynamic data structures. This approach is conceptually straightforward and provides a programmer with a high degree of control over the system. The organization of the system also provides many opportunities for optimizing data access and migration to hide communication latency. The available techniques include using direct memory pointers for accessing data objects, prefetching, protocol tuning, invalidation versus updating for consistency management, optimistic locking, "tagged" objects, and control of many lower-level issues and trade-offs.

The basic mechanisms of the system also provide a suitable foundation for higher-level mechanisms. These mechanisms provide convenient and efficient matrix processing, linked-structure processing, and programming-language interfaces. Techniques for making the best use of the system and translating programs from other distributed-shared-data systems are also discussed.

Obtaining good performance from distributed-shared-data systems is also important. Several example applications are used to demonstrate the performance of the system. As well, the examples provide an opportunity to discuss techniques that are useful in tuning applications for the environment.