Performance Modeling of Large-Scale Parallel Systems and Applications

Adolfy Hoisie, Los Alamos National Laboratory

In this talk we will survey the state-of-the-art in performance analysis and modeling of parallel systems and applications. We will also present novel techniques for performance modeling developed at Los Alamos, including a variety of applications of modeling.

The research space under consideration is multidimensional. It involves analysis of architectures and applications that evolve substantially and rapidly. As a result, modeling and simulation is needed for analysis and predictive capabilities. In this work, the application research space is defined by a scientific workload. The architectural space is composed by current and future large-scale parallel architectures.

We will describe novel methodologies for performance analysis, modeling and prediction applicable to extreme-scale parallel architectures and applications. We will present ongoing and planned future performance projects in PAL at Los Alamos National Laboratory, based on the new methods that can be applied to next generation parallel systems -- 100 T-Ops and beyond.

In addition to the methodology for performance modeling, the applications of the performance models will be presented, including architectural design, workload characterization, application optimization and system diagnostics.

We will conclude by summarizing a number of factors that in our view will significantly impact the development and performance of future generation parallel systems.

This is joint work with Darren Kerbyson.

Created: 9/29/03
Modified: 9/29/03