A science of performance, that is understanding, modeling and predicting the performance of large-scale applications on HPC systems, is one of the great challenges for computer science. Why is performance modeling so difficult? The performance of a particular application on a given machine is a complex function of many variables, and the performance behavioral space is highly non-convex, with many local maxima and minima. To achievea goal of a science of performance, we require scalable modeling methodologies that can be used to characterize performance for a diverse set of applications and parallel architectures.
This poster focuses on techniques that break up the modeling task into manageable parts; we explore the cost-accuracy tradeoffs of several modeling methods. We examine several methods of varying complexity that first separate, and then convolve, metrics of machine and application. The outcome of these investigations is the ability to predict and explain the performance and performance differences of many large-scale HPC applications on many parallel HPC architectures.