<span class="var-sub_title">Purpose-Built HPC: Last Hope for Earth System Prediction?</span> SC18 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Purpose-Built HPC: Last Hope for Earth System Prediction?


Authors: Dave McCarren (US Navy, Oceanographer of the Navy), Mark Govett (National Oceanic and Atmospheric Administration)

Abstract: As earth system prediction models continue to increase in complexity, so do the high-performance computing (HPC) requirements for running these models. Unfortunately, current HPC platforms are designed, built and optimized for a broad array of computationally intensive applications which results in them being inefficient for earth system prediction applications. Specifically, recent top HPC platforms are mostly optimized for machine learning, artificial intelligence, and n-body simulations. They are not designed with the specific purpose of solving partial differential equations (PDEs), such as the Navier-Stokes equations, which are the foundation of earth system modeling.

Long Description: As earth system prediction models continue to increase in complexity, so do the high-performance computing (HPC) requirements for running these models. Unfortunately, current HPC platforms are designed, built and optimized for a broad array of computationally intensive applications which results in them being inefficient for earth system prediction applications. Specifically, recent top HPC platforms are mostly optimized for machine learning, artificial intelligence and n-body simulations. They are not designed with the specific purpose of solving partial differential equations (PDEs), such as the Navier-Stokes equations, which are the foundation of earth system modeling. While current architectures and cloud platforms are suitable for certain computational tasks such as post-processing, their relative inefficiencies for earth system modeling problems will only increase as the commercial markets continue to move toward commodity based solutions. The requirements of supercomputing for earth system modeling are multi-dimensioned; not only floating point operations but also memory system performance and inner-connect bandwidth are crucial for timely execution of earth system simulations.

The HPC working group of the interagency partnership National Earth System Prediction Capability would like to conduct a study investigating the widening gap between application requirements and currently evolving HPC compared with a hypothetical supercomputer designed with the singular purpose of running exascale earth system prediction models. The scope of the study will include: Performance measurement and modeling to characterize quantitatively and in detail computing requirements systematically collected from the earth system modeling community; Correspondingly detailed measurement and characterization of current and roadmap technologies for processor, memory system and network technologies; Gap analysis to determine if custom, including the level of customization,-designed and manufactured PDE-solving processors, interconnect, memory, and other essential components can be created, cost effectively, if needed; Determine if a PDE-solving supercomputing platform would require a specific (and custom) software such as compilers and libraries; Estimation of a rough order of magnitude of investment needed for such a custom-built supercomputer.

Organizing such a study around the what-ifs of such a project will yield valuable insight into the advancing computing platforms for the earth system sciences whether or not a purpose-built earth system prediction computer is actually proposed and built. Ongoing scientific advancements, including increased spatial resolution, earth system component coupling, and data assimilation, will continue to increase the burden on HPC resources in their current state. This study will help identify the current needs of earth system prediction models and how HPC hardware can be built and optimized to alleviate the computational burden moving forward.

We propose a community Birds of a Feather session and panel at the Supercomputing 18 to more fully develop the objectives, scope, methodology, necessary expertise, and likely participants needed for such a study. We are seeking a statement of institutional commitment and support to focus, impel and lend credibility to this proposed study.


URL: http://www.earthsystemprediction.gov/



Back to Birds of a Feather Archive Listing