ExaBounds—Better-than-back-of-the-envelope Analysis for Large-Scale Computing Systems

Large-scale computing systems such as supercomputers and commercial data centers pose many unique challenges in terms of power delivery and heat removal, performance at-scale, monetary cost of purchase and operation, and reliability during operation. Architectures to address these challenges must make appropriate hardware-level choices, in the context of the demands of target applications. Given the huge combinatorial space of choices of algorithms for solving various computational problems, hardware architectures, and technology for system implementation, it is necessary to have a mechanism to identify subsets of the algorithm, architecture, and technology design space worth more detailed study.
EXABOUNDS is an analytic framework being developed for efficiently estimating coarse-grained bounds on, and growth rates of, compute performance, power dissipation, monetary cost, and reliability, of large-scale computing systems. While not intended to enable precise performance prediction as detailed processor or interconnect simulators do, it enables insight into the interaction between performance, power, cost, and reliability, by providing a meaningful yet simple model with complete visibility into the causal relations between system parameters and resulting system behavior. The framework incorporates a large body of empirical technology data, and its utility is demonstrated with a design study for a real future large-scale computing system.