The major goal of this project is to improve the programmability of highly parallel scalable multiprocessing systems (HMPs). PF, taking into account the requirements posed by advanced application codes fro m industry and science. The Meiko CS-2 system at the Vienna Centre for Parallel Computing (VCPC) will serve as the project platform.

The objectives to be achieved in the project are as follows:

Develop project benchmarks
The project benchmarks will be based on a set of advanced application codes, including PAM-CRASH from ESI, IFS from ECMWF and FIRE from AVL.

They will range from highly simplified kernels to representative benchmark programs for requirement specification and evaluation of HPF+. Three increasingly complex versions of the benchmark codes will be developed:

• Version 1 (PM6): simplified kernels of the application codes emulating typical single tasks occurring in the numerical algorithms

• Version 2 (PM15): reconfiguration of simple benchmarks into more complete systems

• Final version (PM21): complete HPF+ version of the project benchmarks

Develop HPF+ language specification
A full specification of syntax and semantics of HPF+ will be developed. HPF+ will eliminate some ill-defined or inefficient features from HPF-1 and include new features according to the requirements of the project benchmarks. A close cooperation with the HPF-2 effort will guarantee that HPF+ will not divert from the standard except where this is necessary for providing functionality not being considered under HPF-2.

• HPF+ Rel.1 (PM4) will extend the data distribution mechanisms of HPF-1 by general block distributions, indirect distributions and data distribution to processor subsets and will provide a mechanism for specifying work distributions in the context of INDEPENDENT loops.

• HPF+ Rel.2 (PM13) will in addition provide processor views, distribution of components of derived types, and mechanisms for controlling dynamic data distribution and the transfer of distributed arguments across procedure boundaries.

• The HPF+ Final Language Definition (PM 23) will in addition provide mechanisms for specifying user-defined distributions and alignments, language features for expressing task parallelism and its integration with data parallelism.

Additional language modifications will be considered during the course of the project, taking into account the analysis of the benchmarks and experiences with the implementation.

Develop HPF+ related compilation technology
The HPF+ extensions will be implemented in the research prototype compiler VFCS. In addition, current VFCS, which is largely based on FORTRAN 77, will be generalized towards Fortran 90. Optimizations will be developed within this framework, in accordance with experiences from using the project benchmarks.

• VFCS extension Vers.1 (PM 9) will include general block and indirect distributions, distribution to processor subsets, independent loops with indirect accesses and explicit work distribution, and support for workspaces.

• VFCS extension Vers.2 (PM 17) will include dynamic data distribution and related interprocedural analysis techniques, support for processor arrays with symbolic bounds, distribution of derived types and dynamic memory allocation.

• VFCS extension Vers.3 (PM 23) will support parallelization of full HPF+.

Extension of MEDEA for performance analysis of HPF+ codes
MEDEA will be extended in order to provide detailed descriptions of the behavior of HPF+ programs and to aid users in the interpretation of the achieved performance. These performance studies will identify problem areas in this implementation, providing feedback to application designers and implementers. In particular the following extensions of MEDEA will be performed:

• Extensions for the automatic analysis of PICL-like tracefiles (PM 7).

• Enhancement of visualization facilities (PM 15).

• Development and integration of a functional description module (PM 23).

Evaluation of HPF+
An evaluation of HPF+ and its implementation, based on the project benchmarks, will provide a comparison of the enhanced technology with HPF-1 and the message passing approach. This evaluation will be carried out with respect to both the required porting effort as well as the performance of the object code at the project platform. The HPF-1 related part of the evaluation effort will use commercial compilers including the HPF Mapper from NA Software. For all kernel migrations to HPF+, the possible impact on the full code will be analyzed. The major steps of the evaluation effort are as follows:

• Evaluation of benchmark programs Version 1 and VFCS extension Version 1 (PM 12).

• Evaluation of benchmark programs Version 2 and VFCS extension Version 2 (PM 18).

• Evaluation of Final Version of benchmark programs and VFCS extension Version 3 (PM 24).

Dissemination and transfer of technology
The project outputs will be widely disseminated. A transfer of technology to industry will be addressed by studying the incorporation of the most effective of the new language features and compilation techniques into commercial compilers such as the NA Software HPF Mapper. Important milestones are:

• Organization of an International Workshop on HPF+ (PM 12).
• Organization of an International Workshop on HPF+ and Related Compilation Technology (PM 24).
• Study of HPF+ and the VFCS extension: classification of language elements and implementation techniques with respect to inclusion into commercial compilers (PM 24).

---