Previous: Conclusions Up: Workshop and Conference... Next: Grand Challenge Requirements Document

Findings

There is nothing quite so frustrating for the avid do-it-yourselfer than to begin a project, suddenly need a particular tool, but have no idea where in the house to look for it.
-- Practical Homeowner's 1987 Do-It-Yourself Annual

Many of the researchers at the Grand Challenge workshop expressed this feeling of frustration in discussing their attempts to find a suitable tool (whether a debugger or a performance evaluation tool), to use a (not-well-documented) tool efficiently, and to reuse it (i.e., have access to it through a software library) in other applications. The workshop participants suggested several new mechanisms for improving the effectiveness of the HPCC program. Indeed, a majority of the suggestions at the workshop focused on the problem of development, maintenance, and distribution of tools. Here we summarize three of the principal suggestions and-more important-propose a mechanism for implementing each suggestion.

Suggestion: Collaborative development of tools for Grand Challenge applications should be supported. Special attention should be given to parallel debuggers, performance evaluation tools, and interactive visualization software. Such efforts must be supported for all phases of the life cycle-fundamental research, prototyping, and development.
Mechanism: Support a ``Grand Challenge in Software Tools," and establish a Science and Technology Center for HPCC Software Tools.
Suggestion: Tools that are ready for use outside of their development community should be made available in software-sharing libraries.
Mechanism: Establish or fund existing ``software capitalization" programs to foster the development of research ideas and prototypes into production tools and to ensure the maintenance of the software in the library.
Suggestion: Tools that have had success in the research community should be commercialized.
Mechanism: Fund an existing organization such as OSF, the national laboratories, or the NSF supercomputer centers; or set up a new independent organization or industrial consortium.
Suggestion: Researchers need to be kept informed of the numerous tools that are available. At the same time, computer scientists must be made aware of the needs of applications scientists for specific tools.
Mechanism: Develop a parallel software tools electronic journal/newsletter (complete with an editorial board) to publicize tool availability and tool needs.

That the computer scientist is caught between Charybdis and Scylla-encouraged to create prototype tools as research projects, but given little credit for developing production-quality tools for widespread use-was clearly recognized by the conference participants. Many felt that the vendors must assume more responsibility in providing useful tools.

Suggestion: The quality and performance of the basic HPCC tools should be improved. In particular, vendors should provide good node compilers to allow programmers to make more effective use of machines without resorting to assembly language coding; provide fast multiple-precision arithmetic; pay greater attention to symbolic and combinatorial aspects of computing, such as fixed-point arithmetic; provide language constructs that allow the easy definition of complex data structures; provide high-level, standardized message-passing routines; provide a random number generator with their machines; and develop a compiler infrastructure for making visible the compiler's internal representations and transformations to other software tools.
Mechanism: Require computer vendors (through explicitly worded RFPs) to provide good, usable tools for MPP machines. Have federal agencies support commercial development of software tools.

Of course, providing appropriate tools (whether from vendors or from computer scientists) requires an awareness of what is needed. While some applications projects explicitly made provision for ``computer science work,'' most do not. The workshop participants agreed that much more interaction is needed.

Suggestion: New incentives should be developed to encourage computer scientists and application scientists to interact more closely on Grand Challenge projects.
Mechanism: Have HPCC projects explicitly fund the exchange of students and staff between application and tool developers.

Software technology requirements certainly extend beyond tools for Grand Challenge applications research. Several areas were identified at the workshop as needing increased attention. Here we first list the various suggestions; a single mechanism is proposed to address these suggestions.

Suggestion: High-performance file systems must be developed to ensure ubiquitous access to both conventional and high-performance distributed file systems.
Suggestion: Database management systems must be developed that (1) support large objects, (2) provide fast access to subsets of large datasets, (3) interface to tertiary storage systems, and (4) incorporate new indexing and access methods for scientific data.
Suggestion: Methods should be developed to allow mixing of different languages in a single framework.
Suggestion: ``Templates" that capture key computational, communication, and I/O characteristics of important application codes should be written and made widely available.
Suggestion: More emphasis should be placed on integrating support for I/O and networked services.
Suggestion: Standards efforts such as the Message Passing Interface forum should be encouraged.
Suggestion: Heterogeneous computing must be fully funded through large projects that include software, applications, and network research.
Suggestion: Engineers and scientists must be retrained in multidisciplinary technologies, and new researchers must be educated in this critical HPCC area through the establishment of multidisciplinary curricula at universities.
Mechanism: Increase the current federal involvement, specifically by providing (1) stable and long-term funding for new initiatives and (2) stable base-level funding for ongoing projects.

stevens@mcs.anl.gov