Architecture Simulator UQ

Sensitivity and Uncertainty Analysis of a Macroscale Architecture Simulator

Computer architecture simulators are a key technology in the design and performance analysis of high-performance computing (HPC) systems. Sandia has a suite of simulators aimed at multiscale simulations of HPC platforms ranging from clock-accurate models up to coarse-grained simulation of application behavior at the exascale. Our group is developing methods to assess the predictive fidelity and analyze sensitivity in macroscale architecture simulators aimed at enabling hardware/software co-design for extreme scale parallel systems.

Co-design presents significant engineering challenges associated with jointly assessing the impact of engineering parameters for proposed computing platforms and software design decisions for the applications of interest. For instance, the presence of both discrete and continuous design parameters, discontinuities in the model output space, and (often unknown) correlations between the input parameters challenge the existing approaches. We develop advanced uncertainty quantification and propagation methods to address these challenges. The tools will be used, for example, to identify key design parameters and to analyze the robustness of the performance of proposed computer architectures with respect to a variety of programming models and applications.