The project implements a virtual-memory system for data-parallel computing by taking advantage of the technical advances in algorithms for parallel disk systems. Such a system will benefit parallel computing in many scientific and other application areas by making it simpler and faster to run code on objects that are several gigabytes or even larger. Typically, when the memory requirements of an application exceed the available random-access memory on parallel machines, programmers keep their data on a parallel disk system and perform disk accesses in their source code. The resulting codes are difficult to develop and often run slowly. The proposed system will help avoid the additional software development time and complexity associated with such large problems. Moreover, because the system employs recent advances in algorithms for planning parallel disk accesses, it runs more efficiently. The project entails the design and implementation of an appropriate set of primitive data-parallel operations in a virtual-memory context. The implementation may include further development of efficient algorithms for parallel disk systems. In addition, the project includes the development of one or more compilers that compile languages such as Fortran and C into these primitives.
