Performance Models | Algorithms & Applications Group

Performance: Monitoring, Modeling, and Predicting

Performance: Monitoring, Modeling, and Predicting
supported by NSF, Dept of Education, and NATO
Mark Mathis, Jack Perdue
Nancy Amato, Andrea Pietracaprina, Geppino Pucci

The overall goal of our research is to develop techniques and the infrastructure to collect, manage and utilize performance metrics and to use them to develop techniques for automatic parallel performance prediction. We hope the contribution of this work will be a methodology by which accurate predictions of parallel performance can be made that will assist the application and user to better utilize the computing resources available. An important aspect of our work is to develop techniques that can be used at run-time in order to customize performance for the current application (program and data) and the current system conditions. We will consider a variety of approaches that range from using little or no application-specific knowledge to methods that require a detailed analysis of the specific algorithm.

As part of this work, we will integrate our metric collection and analysis framework into STAPL, the Standard Template Adaptive Parallel Library. An important aspect of STAPL is the STAPL runtime system which will provide a mechanism to handle communications in distributed memory and synchronizations in shared memory systems and dynamically repartition computations based on current system architecture and utilization. It is also responsible for choosing algorithmic options that are best suited for the current circumstances (e.g., what type of parallel sort to do). The ability to make accurate predictions based on different partitioning schemes (e.g., data and computation), current system utilization (e.g., number and type of jobs presently running) and algorithmic options/choices is fundamental to its runtime decision making. One of the primary focuses of this work is to provide the performance predictions needed by the STAPL runtime to utilize the parallel system to its fullest.

Papers

A General Performance Model for Parallel Sweeps on Orthogonal Grids for Particle Transport Calculations, Mark M. Mathis, Nancy M. Amato, Marvin Adams, In Proc. ACM Int. Conf. Supercomputing (ICS), pp. 255-263, Santa Fe, NM, May 2000. Also, Technical Report, TR00-004, Parasol Laboratory, Department of Computer Science, Texas A&M University, Dec 1999.
Proceedings(ps, pdf, abstract) Technical Report(ps, pdf, abstract)

Predicting Performance on SMPs. A Case Study: The SGI Power Challenge, Nancy M. Amato, Jack Perdue, Andrea Pietracaprina, Geppino Pucci, Mark Mathis, In Proc. Int. Par. and Dist. Proc. Symp. (IPDPS), pp. 729-737, Cancun, Mexico, May 2000. Also, Technical Report, TR99-020, Department of Computer Science and Engineering, Texas A&M University, Oct 1999.
Proceedings(ps, pdf, abstract) Technical Report(ps, pdf, abstract)

Comparing the Memory System Performance of the HP V-Class and SGI Origin 2000 Multiprocessors using Microbenchmarks and Scientific Applications, Ravi Iyer, Nancy M. Amato, Lawrence Rauchwerger, Laxmi Bhuyan, In Proc. ACM Int. Conf. Supercomputing (ICS), pp. 339-347, Rhodes, Greece, Jun 1999.
Proceedings(ps, pdf, abstract)

A Cost Model for Communication on a Symmetric MultiProcessor, Nancy M. Amato, Andrea Pietracaprina, Geppino Pucci, Lucia K. Dale, Jack Perdue, Technical Report, TR98-004, Department of Computer Science and Engineering, Texas A&M University, Presented at SPAA Revue, 1998., Jan 1998.
Technical Report(ps, pdf, abstract)

Dissertations and Theses

A General Performance Model for Parallel Sweeps on Orthogonal Grids for Particle Transport Calculations, Mark M. Mathis, Masters Thesis, Department of Computer Science and Engineering, Texas A&M University, Dec 2000.
Masters Thesis(ps, pdf, abstract)

Developing A Cost Model for Communication on a Symmetric MultiProcessor, John Kimbal Perdue, Masters Thesis, Department of Computer Science and Engineering, Texas A&M University, Dec 1998.
Masters Thesis(ps, pdf, abstract)