Performance models are an important tool for gaining insight into applications and systems. They can be used for scalability analysis on both existing and proposed future architectures, in procurement to compare proposed alternatives, in software development to ascertain the performance impact of code re-configuration prior to implementation, and in real-time to steer the processing of code to increase processing efficiency. Accurate models are useful for guiding architecture design. The modeling of high performance software and hardware systems is highly complex, requiring the encapsulation of key processing structures and characteristics. This is a direct result of the performance space being multi-dimensional and highly non-linear in any of its dimensions.

Research in this area at LANL and Rice will span a wide range of topics. At Rice, the focus of research will be on designing, building and evaluating semi-automatic tools for synthesizing models and model components, as well as exploring how to integrate model components synthesized automatically into hand-crafted model frameworks. We will use the tools that we build to synthesize prototype models for target applications and then evaluate how effectively our semi-automatically generated models predict application behavior on a set of target systems.