Retire Fortran?: a debate rekindled

A Debate Rekindled-of 1984, te between ~cGraw ~awrence ivermore tional Laboratory (LLNL) and David Kuck and Michael Wolfe of Kuck and Associates, appeared in Physics Today [19]. The subject was whether to retire Fortran. Eight years have passed, and we wish to reopen the debate and provide further evidence that Fortran is not the sine qua non of high-speed computing. Many believe that the existing investments in Fortran and the quality of existing Fortran compilers are preventing a change in programming methodology. Many also feel that support for Fortran must continue because the language is familiar and widely available. Unfortunately , the complexity of writing correct parallel programs in Fortran is perpetuating today's software crisis. We believe, as did McGraw in 1984, that increased productivity, generality, utility, portability, and performance are only possible if programmers avoid the constraints of imperative languages and adopt a higher level of abstraction. We must escape the morass of imperative semantics and attain a level of abstraction that separates the programmer from the machine, stresses problem definition over the mechanics of solution, and provides complete information to the compiler. While Kuck and Wolfe would probably have agreed with the above statements, they wisely argued that programmers would not use languages that did not allow them to get the performance they required. To this McGraw had no counterargument, for in 1984, only Fortran provided the performance needed for large-scale scientific computing. The intent of this article is to reopen the debate and present overwhelming evidence that functional supercomputing is possible. To this end, we compare the execution performance of Sisal, a functional language for large-scale scientific computing, and For-tran on a Cray Y-MP/864. The Crays remain the most heavily used machines in our nation's supercom-puter centers, and are the preferred machines of most scientific programmers. Thus, it is here that languages must outperform For-tran if they are to replace it. For completeness and for the benefit of those readers not familiar with the principles of functional computing, the next two sections examine the functional programming paradigm, discuss its attributes and advantages, and highlight the salient features of Sisal. In the remaining sections we illustrate the potential inefficiencies of functional computing, present our most recent performance data, and give some closing remarks regarding Fortran and the future of high-speed computing. In 1984, McGraw noted that, by all indications, future supercomputers would be multiprocessors. Today, most supercomputer users …