Programming Parallel Supercomputers

This paper discusses future directions in tools and techniques for programming parallel supercomputers We base the discussion on two important observations Automatic parallelization of sequential programs will not achieve supercomputer performance in re al applications Instead applications will have to be written with explicit parallelism Machine independence of parallel programs is a precondition for wide acceptance of parallel com puters We comment on High Performance Fortran HPF and conclude that HPF will achieve machine independence to an initially satisfactory degree but that another language revision can be expected Machine independence does not imply poor perfor mance We present evidence that explicitly parallel machine independent and problem oriented programs can be translated automatically into parallel machine code that is competitive in performance with hand written code Furthermore we show that interactive source level and problem oriented debugging of explicitly parallel program has recently become a reality