NHSE Review 1996 Volume Second Issue: Comparison of 3 Compilers -- Chapter 1

NHSE Review^TM 1996 Volume Second Issue

Comparison of 3 HPF Compilers

| <- HREF="author-and-abstract.html" Prev | Index | Next -> |
NHSE Review^TM: Comments · Archive · Search

Chapter 1 -- Three HPF Compilers for the IBM SP

Three compilers for HPF that target the IBM Scalable POWERparallel System (SP) and networked clusters of IBM RISC System/6000 workstations running AIX are reviewed and contrasted:

Applied Parallel Research, Inc., xHPF, Version 2.1(11), available to all CTC users since mid-July 1996
Portland Group, Inc., pghpf, Version 2.1, available to all CTC users since mid-July 1996
IBM, XL HPF, Version 1.1.0.0, available to all CTC users since April 19, 1996

(The noted dates are for actual release of the product version of each compiler at that revision-level to all users of the Cornell Theory Center IBM SP complex. The APR and PGI products were available in the field at other sites at much earlier dates -- APR's product line was in use at CTC from about mid-1992 -- and the IBM product was in Beta Test at CTC before April 1996, with a limited number of researcher-users of CTC employing earlier non-product versions in their computational research.

Note also that there are other HPF compilers available for the IBM SP, such as VAST-HPF from Pacific-Sierra Research, and ADAPTOR from the Institute for Algorithms and Scientific Computing (SCAI) at the German National Research Center for Information Technology. These have not been reviewed for this paper.)

1.1 Summary of Conclusions

There exist effective, competing HPF products for the SP.

There are significant differences between all three compilers.

All handle significant amounts of the HPF 1.1 Subset language and have extensions into the full HPF 1.1 language [1]. (pghpf handles most of full HPF 1.1)
All handle Fortran 90 at least in part. (pghpf and XL HPF handle most of full Fortran 90 in programs that are Fortran 90 and HPF.)
Each has its strengths for particular styles of program source.
For parallel execution tuning, each provides a different set of facilities and diagnostics.
If budget permits, acquiring more than one of these is worthwhile.

1.2 Introduction

This paper is an outgrowth of a number of talks, presentations, and training modules that the author prepared or participated in over the past five years at the Cornell Theory Center (CTC). During this time, the development of HPF as a standardized language for data parallel Fortran programming of distributed memory computers was first supported by the CTC. Then as HPF processors became generally available for the IBM SP, some were acquired and evaluated, documentation was augmented as required, training of staff and users was performed by the CTC, and various levels of user consulting was done. (It should be noted that none of the evaluations were organized as rigorous benchmarking efforts.) Over this period, the author has compared and contrasted the generated-code performance of these compilers, in conjunction with his Parallel Tools work with CTC colleagues.

The overall aim of this paper is to describe these compilation systems and to contrast the features supported by each compiler, presenting them in a compact form for those interested in either using a particular HPF product on an IBM SP, or in acquiring HPF resources for an IBM SP.

This paper assumes that the reader is familiar with FORTRAN 77, Fortran 90, and High Performance Fortran (HPF). Good online training material on these topics can be found at:

The High Performance Fortran standard (Version 1.1) can be found online at:

: High Performance Fortran Forum, High Performance Fortran Language Specification, Version 1.1

The author is particularly in debt to Donna Bergmark, the project leader of the CTC Parallel Tools Team, who first promoted the furtherance and the use of HPF at the CTC. Many of the example codes in this paper are the result of her work or of collaborations with which she has been a partner.

1.3 Organization of this Report

Chapter 2 contains the main material of the evaluation. After a brief discussion of the evaluation criteria, four tabularizations are presented:

HPF Language Features
Other Capabilities (parallelization, Fortran 90 coverage, etc.)
Tools and Extensions
Subjective Evaluations

It must be noted at the outset that this paper is not the result of a "benchmarking" effort, either of the compilation-time performance of the three compilers or of the generated target code.

Chapter 3 describes the architecture of each compiler at a high level and gives some general information about the run time system to support parallel execution that each uses.

Chapters 4, 5, and 6 exhibit details of code generated by each compiler, contrasting between them, for small HPF codes. Complete listings, in intermediate FORTRAN 77 or in a pseudo-Fortran form are available for each example.

The only conclusions drawn by this report are the comparisons summarized in Chapter 2 and thus there is no separate "Conclusions" chapter.