Performance Engineering Research

Enhancing the performance of SciDAC applications on petascale systems

Robert F. Lucas (project website)
University of Southern California

While current terascale and planned petascale supercomputers offer unprecedented capabilities for creating detailed scientific simulations and analyzing massive amounts of data, making the most efficient and effective use of such systems requires that both the systems and the applications be optimized for highest performance. To help achieve this goal, the Performance Engineering Research Institute will work with SciDAC applications teams and the research community through tutorials and workshops on improving application performance (typically associated with major conferences). These activities will help disseminate results and maximize impact on the performance of SciDAC and other high priority DOE scientific codes.

The Institute will also collaborate with the other SciDAC Centers and Institutes and the DOE scientific computing facilities to enhance the performance of a broad set of application projects. Each year, the Institute will target a particular code or set of codes within a given discipline for performance analysis and optimization to achieve specific short-term performance goals. The DOE scientific computer facilities will also work with this Institute to add in-depth questions to their regular user surveys to solicit users’ experience with performance problems and tools and to collect data on the use of performance tools. The Institute will also conduct surveys to solicit information from application developers about current application performance and future performance goals. Experience has shown that short-term intensive workshops can be effective in overcoming the tool learning curve, obtaining initial results, developing collaborative working relationships, and determining future directions for productive performance optimization efforts. Therefore, workshops, tutorials and more formal short courses or summer courses will also be developed.

In the future, achieving good performance on petascale computing systems will grow ever more challenging due to enormous scale and increasing complexity in both architecture and application. Further, users want to focus on their science, and not be burdened by the need to optimize code performance. Thus, the ideal performance tool analyzes and tunes performance automatically. This envisions tools that analyze a scientific application, both as source code and during execution, generate a space of tuning options, and search for a near-optimal performance solution.

There are many challenges to realizing this vision, including enhancement of automatic code manipulation tools, automatic run-time parameter selection, automatic communication optimization, and intelligent heuristics to control the combinatorial explosion of tuning possibilities. To address these challenges, the Performance Engineering Research Institute will focus on:

  • education and outreach
  • performance modeling and prediction
  • automatic performance optimization, and
  • performance engineering of high profile applications.

A performance modeling and prediction activity will develop and refine performance models, significantly reducing the cost of collecting the data upon which the models are based and increasing model fidelity, speed and generality. A new component research activity in automatic performance optimization tuning software is spurred by the strong user preference for automatic tools. Performance engineering of high-profile applications will direct interactions with SciDAC applications. To ensure that important performance goals are achieved in the near term, the Institute will assign personnel and allocate resources to work directly with individual science application projects as needed.

SciDAC Institute: Computer Science

Project Title: Performance Engineering Research Institute

Principal Investigator: Robert F. Lucas
Affiliation: University of Southern California

Project Website:

Participating Institutions and Co-Investigators:
Argonne National Laboratory - Paul Hovland, Dinesh Kaushik, Boyana Norris
Lawrence Berkeley National Laboratory - David Bailey, Daniel Gunter, Katherine Yelick
Lawrence Livermore National Laboratory - Bronis de Supinski, Daniel Quinlan
Oak Ridge National Laboratory - Jeffrey Vetter, Patrick Worley
Rice University - John Mellor-Crummey
University of California at San Diego - Laura Carrington, Allan Snavely
University of Maryland - Jeffrey Hollingsworth
University of North Carolina - Robert Fowler, Daniel Reed
University of Southern California - Jacqueline Chame, Mary Hall, Robert F. Lucas (PI)
University of Tennessee - Jack Dongarra, Shirley Moore

Funding Partners: Office of ScienceOffice of Advanced Scientific Computing Research

Budget and Duration: Approximately $3.0 million per year for five years 1

Other SciDAC Institutes
Other SciDAC computer science efforts

1Subject to acceptable progress review and the availability of appropriated funds


Home  |  ASCR  |  Contact Us  |  DOE disclaimer