In the very broadest of terms, our work is basic research on problems that need to be solved to provide effective, mainstream parallel computation across a wide range of architectures, from networked desktop boxes to parallel and vector supercomputers. For parallel computing to become mainstream will require programming models that are high-level and machine-independent. Our group focus is therefore on fundamental issues in parallelizing compiler technology. Specifically, we are working on (1) very aggressive parallelization and optimization, (2) generalizing our optimization and code generation techniques to target architectures spanning message-passing systems (e.g. MPI), hardware-shared memory (both uniform and NUMA/DSM), software DSMs (e.g. TreadMarks and Brazos), distributed object memory systems, and hybrids combining hardware shared memory with other communication architectures, (3) integrating performance debugging tools with the compiler so that application programmers can effectively work in the high-level language, and (4) compiling programs with data sets too large for main memory. Our research infrastructure is the dHPF compiler (https://www.crpc.rice.edu/fortran-tools/DSystem/dhpf/).

The current use of Windows NT within our group per se is limited to traditional desktop tasks and some scientific visualization applications. In the near term, however, we are committed use various configurations of NT systems and clusters as target architectures for our compiler research. This is being driven from several directions. Large parts of the application community are interested in reducing the cost of parallel computing through the use of commodity (NT on Intel) systems, so we are investigating the issues of compiling directly for clusters of multiprocessors. Our favorite software DSMs for LANs were either designed for NT (Brazos) or are being ported to it (TreadMarks). We are partners with both NCSA and NPACI; through these relationships we are committed to geographically-distributed high-performance computing and at least some of this will be done on NT-based systems. In the medium term this work will entail hosting our programming tools on NT. In the longer term it may involve research in parallelizing compilers for such environments.

Robert J. Fowler
Department of Computer Science
and Center for Research on Parallel Computation
Rice University