This research deals with issues related to the efficient use of parallel I/O systems. These include problems of data placement, caching, and prefetching and their interaction in determining system performance. The project involves the development of robust parallel I/O models that incorporate these requirements, and designing new algorithms to obtain desired performance levels. This includes identifying the appropriate policies for resource management in the context of both single and multiple interacting computational and I/O streams, and the development of suitable algorithmic strategies. The focus is on on-line buffer management and prefetching algorithms with bounded lookahead. The use of randomization in data placement and replacement, and issues arising in their implementation are under consideration. Theoretical analysis of the algorithms is based on a competitive framework to compare different on-line methods, and identify promising candidates. Suitable candidates will be implemented and their performance validated empirically. The areas of impact include, though are not restricted to, data-intensive applications like modern database and multimedia systems.
