Keynote Address: The Prospects For Architecture-independent Parallel Programming

Parallel computing has not lived up t o its promise. After nearly a decade of research, parallel systems are found primarily in research labs and only a few corporations have adopted it for their mainline computing problems. Significantly, very few independent software vendors have ported their codes t o parallel machines. Now high-performance workstations are rapidly gaining on high-end supercomputing systems and we are undergoing a shakeout among the vendors of parallel computer systems. Is parallel computing dead? In this address, I will argue that i t is premature t o declare its demise, but that key software problems need to be addressed if it is t o succeed. Foremost among these is the problem of support for architecture-independent parallel programming. Software developers need t o be able t o rely on the retargetability of their codes for new parallel machines as they emerge. The absence of such assurance has been the prime reason for the limited acceptance of scalable parallelism to date. As an example, I will discuss High Performance Fortran, which is designed to support the construction of architecture-independent data-parallel programs. High Performance Fortran extends Fortran 90 by providing a set of statements that specify the distribution of data structures across the processor array. From this specification, the compiler automatically derives parallelism according to the “owner computes” rule, which specifies that a computation should be performed on a processor that owns most of the data involved in the computation. Although over twenty companies are working on products related t o High Performance Fortran, its success is not yet guaranteed. To be widely accepted, High Performance Fortran will need a very sophisticated compiler technology. In addition, it needs additional features to increase the breadth of coverage t o include irregular problems, task parallelism, and high-performance input/output. I will survey the current strategies for meeting these needs and describe the tools needed to make programming in Fortran D accessible t o the average scientist or engineer. The talk will conclude with a discussion of the long-term prospects for scalable parallel computation and the software needed t o make it practical.