Sizing equivalent performing large-scale systems using modeling

• Published in: Innovative Architecture for Future Generation High-Performance Processors and Systems, 2003 pp. 46 - 55
• Main Authors: Kerbyson, D.J., Hoisie, A.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 2003
• Subjects: Analytical models; Computational modeling; Earth; Hardware; Laboratories; Large-scale systems; Microprocessors; Performance analysis; Predictive models; Procurement
• ISBN: 0769520197; 9780769520193
• ISSN: 1537-3223; 2642-4118
• DOI: 10.1109/IWIA.2003.1262782

This work analyses the relative performance of the Earth Simulator to systems constructed from HP AlphaServer nodes such as ASCI Q. The Earth Simulator uses vector processing nodes interconnected using a single-stage cross-bar network. The AlphaServer systems use microprocessors with deep memory hierarchies and are typically interconnected using the Quadrics Elan3 fat-tree network. The performance that can be achieved on a system result from an interplay of the characteristics of the system, the requirements of the workload, and scaling behavior of the two. Detailed performance models are used here to predict the performance of two codes representative of ASCI computations, namely SAGE and Sweep3D. The models are used to compare the performance of the two systems and in particular size an AlphaServer system that achieves the same performance as the Earth Simulator on SAGE and Sweep3D.