High Performance Fortran

High Performance Fortran
Name	High Performance Fortran
Paradigm	Procedural, array-oriented, parallel
Developer	Intel Corporation, Los Alore, Cray Research, National Science Foundation
First release	1993
Latest release	1997 (HPF 2.0 proposals)
Influenced by	Fortran 90, Fortran 77, Fortran 95
Influenced	ZPL (programming language), CAF (Coarray Fortran), Fortran 2008
Typing	Static, strong
License	Various proprietary and open-source implementations

High Performance Fortran is an extension of Fortran 90 designed to express data parallelism and guide distributed-memory compilation for scientific and engineering workloads. Developed in the early 1990s by a consortium of industry, government, and academic organizations to target distributed-memory multiprocessors and vector supercomputers, it introduced directives and array distribution features to map array computations onto parallel hardware. HPF influenced subsequent standards and implementations in the supercomputer and high-performance computing communities.

History

HPF originated from collaborative efforts involving Intel Corporation, Cray Research, the National Science Foundation, and academic groups such as University of California, Berkeley and Rice University. The language effort coalesced around the 1992–1993 period with the publication of the HPF language specification and was shaped at gatherings like Supercomputing Conference meetings where vendors such as IBM and Thinking Machines Corporation participated. Funding and research programs from agencies like the Defense Advanced Research Projects Agency and the UK Science and Engineering Research Council supported compiler research and benchmarking. Subsequent revisions and proposals through the late 1990s, including work toward HPF 2.0, interacted with committees for Fortran 95 and influenced later standardization efforts at ISO/IEC JTC 1/SC 22.

Language Features

HPF extended Fortran 90 with directives and intrinsic abstractions: data distribution directives like ALIGN and DISTRIBUTE, the INDEPENDENT directive for loop-level parallelism, and FORALL and PURE semantics influencing array assignment. The language defined templates, processor arrays, and distribution formats (BLOCK, CYCLIC) to target architectures such as systems designed by Cray Research, Connection Machine, and early clusters promoted by Intel Corporation. HPF semantics relied on concepts formulated in academic work at Massachusetts Institute of Technology, Lawrence Livermore National Laboratory, and Los Alamos National Laboratory to balance locality and communication. Influence from projects at University of Illinois at Urbana–Champaign and University of Cambridge shaped the mapping of array operations to distributed-memory machines.

Compiler Implementations and Tools

Major compiler and tool vendors produced HPF implementations: Intel Corporation provided HPF compilers for its product lines; Cray Research integrated HPF features into its Fortran compilers on vector and parallel systems; research compilers emerged from Rice University, University of Houston, and University of Tokyo. Tools for performance tuning and debugging involved collaborations with institutions such as Sandia National Laboratories and Lawrence Berkeley National Laboratory, and commercial offerings from Absoft Corporation and PathScale implemented HPF-like features. Benchmarking suites comparing HPF implementations leveraged testbeds at Argonne National Laboratory and National Center for Supercomputing Applications.

Performance and Parallelism

HPF targeted distributed-memory parallelism common on systems from Cray Research, Thinking Machines Corporation, and emerging commodity clusters using hardware from Intel Corporation and DEC (Digital Equipment Corporation). By specifying array distribution across processor arrays, HPF enabled compilers to perform loop transformations, communication generation, and locality optimizations influenced by research at University of Illinois at Urbana–Champaign and Stanford University. Performance evaluations used benchmarks from NAS (NAS Parallel Benchmarks) communities and studies reported at International Conference for High Performance Computing, Networking, Storage and Analysis and SC Conference forums. HPF’s effectiveness varied by problem class and by sophistication of backend support from vendors like IBM and SGI.

Adoption and Impact

Adoption encompassed national laboratories such as Lawrence Livermore National Laboratory, Los Alamos National Laboratory, Oak Ridge National Laboratory and industrial research centers at Siemens and General Electric. HPF shaped pedagogical and research directions at universities including MIT, Stanford University, and University of California, Berkeley and influenced successor standards like Fortran 2008 and language designs such as CAF (Coarray Fortran) and ZPL (programming language). Industrial uptake was uneven: some supercomputer centers adopted HPF directives for legacy Fortran codes, while others favored message-passing models championed by groups at Argonne National Laboratory and Los Alamos National Laboratory that developed MPI.

Criticism and Limitations

Critics from academia and industry including researchers at University of Illinois at Urbana–Champaign and Rice University highlighted limitations: difficulty of expressing irregular communication patterns, reliance on sophisticated compiler analysis, and inconsistent vendor support across systems from Cray Research, SGI, and IBM. Performance portability concerns were raised at forums such as Supercomputing Conference and by projects funded by the National Science Foundation. HPF’s high-level abstractions sometimes obscured low-level tuning needed on platforms produced by Intel Corporation and DEC (Digital Equipment Corporation), leading many centers to prefer explicit approaches like MPI or vendor-specific extensions.

Category:Fortran