This project describes work in portable, high performance algorithm development, with experimental system support for I/O intensive programs. For these programs the data access time is the predominant component of the running time. The work has three thrusts. (1) Multiprocessor Multi-disk File Systems. In ongoing research, an experimental multiprocessor multi-disk file system has been built, which runs on Intel Paragon. The file system was built to support a prototype for a standard application programmer interface; the interface has been designed to facilitate the production of portable programs. A major portion of this project is to evaluate the efficiency of theoretically optimal data access algorithms that have already been designed, and to validate the paradigms used to design these algorithms for a two-level memory hierarchy. (2) Additional Two-Level Memory Model Algorithms. LAPACK is a natural (and very appropriate) package for upgrading to accommodate for out-of-core data sets. A portion of these algorithms will be redesigned and implemented as prototypes for testing and validation. Completed programs will be released for public use and evaluation. (3) Three-level Memory Model. A significant research goal is to develop a memory which captures critical performance and latency behavior between main memory, disk systems, and tape systems, for many of the contemporary computer architectures. Such a model is intended to support the development of I/O- efficient algorithms.
