Randolph G. Scarborough

A vectorizing Fortran compiler

This paper describes a vectorizing Fortran compiler for the IBM 3090 Vector Facility. Opportunities for vectorization are investigated for eight levels of DO-loop nesting. Recurrences in inner loops do not prevent vectorization of outer loops. A least-cost analysis determines from the opportunities identified which specific vectorization will result in the fastest execution. The normal optimization phases of the compiler produce much of the information needed for the vectorization analysis.

Improved optimization of FORTRAN object programs

For many years the FORTRAN H Extended compiler has produced highly optimized object programs for IBM System/360 and System/370 computers. A study of the object programs revealed, however, that important additional optimizations were possible, and the compiler has been enhanced accordingly. First, the range of cases handled by the optimization techniques already present in the compiler has been extended. For example, more duplicate computations are eliminated, and more invariant computations are moved from inner to outer loops. Second, several new optimizations have been added, with subscript computation and register allocation receiving particular attention. Third, certain optimization restrictions have been removed. This paper describes these improvements and reports their effects.

Clustering IBM Enterprise System/3090 computers for parallel execution of FORTRAN programs

Two IBM Enterprise System/3090TM Model SOOJ computer systems, each with six processors capable of executing vector and scalar instructions, have been connected into a cluster for parallel execution of single FORTRAN programs. The clustering Is achieved through a combination of software and hardware. When enabled for parallel execution and allowed to use all twelve processors in the cluster, FORTRAN programs have run as much as 11.7 times faster than when run on a single processor. The combined hardware and software technology Is called IBM Clustered FORTRAN. It was achieved by modifying existing technology quickly to provide new capabilities. This paper discusses the modifications and the motivations behind them. It summarizes the performance of several applications executed with Clustered FORTRAN. Finally, it describes how clustering has been used to improve performance in novel ways.