Fortran RED - A Retargetable Environment for Automatic Data Layout

The proliferation of parallel platforms over the last ten years has been dramatic. Parallel platforms come in diierent avors, including desk{top multiproces-sor PCs and workstations with a few processors, networks of PCs and workstations, and supercomputers with hundreds of processors or more. This diverse collection of parallel platforms provide not only computing cycles, but other important resources for scientiic computing as well, such as large amounts of main memory and fast I/O capabilities. As a result of the proliferation of parallel platforms, the \typical proole" of a potential user of such systems has changed considerably. The specialist user who has a good understanding of the complexities of the target parallel system has been replaced by a user who is largely unfamiliar with the underlying system characteristics. While the specialist's main concern is peak performance, the non{specialist user may be willing to trade oo performance for ease of programming. Recent languages such as High Performance Fortran (HPF) and SGI Parallel Fortran are a signiicant step towards making parallel platforms truly usable for a broadening user community. However, non-trivial user input is required to produce eecient parallel programs. The main challenge for a user is to understand the performance implications of a speciied data layout, which requires knowledge about issues such as code generation and analysis strategies of the HPF compiler and its node compiler, and the performance characteristics of the target architecture. This paper discusses our preliminary experiences with the design and implementation of Fortran RED , a tool that supports Fortran as a deterministic, sequential programming model on diier-ent parallel target systems. The tool is not part of a compiler. Fortran RED uses HPF as its intermediate program representation since the language is portable across many parallel platforms, and commercial and research HPF compilers are widely available. Fortran RED is able to support diierent target HPF compilers and target architec-tures, and allows multi{dimensional distributions in addition to dynamic remapping. This paper focuses on the discussion of the performance prediction component of the tool and reports preliminary results for a single scientiic kernel on two target systems, namely PGI's and IBM's HPF compilers with IBM's SP{2 as the target architecture.