Hardware Evaluation

Lead: Scott Pakin, Los Alamos National Laboratory

Hardware Evaluation (HE) staff provide independent architectural analyses of exascale hardware components—even those that have not yet materialized—by using a variety of analysis tools and approaches. Their findings inform system design, highlight opportunities and challenges for scientific applications, and provide understanding of observed performance and efficiency from a hardware and software perspective.

Description: Once a supercomputer is installed, configured, made available to users, and benchmarked, it is too late to consider fundamental architectural changes. The benefit of predictive analysis is that it guides US Department of Energy (DOE) supercomputer procurements toward the systems that deliver the most performance to the extreme-scale applications of interest to DOE.

The HE project element establishes a bridge between application developers and hardware vendors. This bridge goes both ways in that HE provides architecturally insightful and application-relevant feedback to vendors and informs computational scientists of important emerging architectural and performance trends so that they are prepared when new DOE supercomputers become available. HE’s specific work falls into two primary categories.

  1. Performance analysis: Gather and answer hardware-related performance questions held by the DOE facilities and/or the Exascale Computing Project (ECP) application development (AD)/software technology (ST).
  2. Tool development: Enhance DOE performance-analysis software tools—almost all of which are open-source—with new features and/or improved software engineering and/or improved documentation.

Having these capabilities within DOE is critical because HE staff understand DOE application needs and can provide objective analyses of hardware features and unbiased comparisons of current and future hardware platforms. To that end, HE comprises three working groups, each with its own expertise: Analytical Modeling and Node Simulation, Memory Technologies, and Interconnect Simulation. Collectively, these working groups can approach a problem from multiple angles and provide DOE facilities and ECP AD/ST with multifaced information on application and system performance.

Lead: Scott Pakin, Los Alamos National Laboratory

National Nuclear Security Administration logo Exascale Computing Project logo small U.S. Department of Energy Office of Science logo