Enhancing Computational Science Opportunities on NSF's Kraken

点击翻译按钮获取中文摘要

As part of the Office of Cyberinfrastructure program to provide advanced computational science resources to the U.S. national computational science community, the Kraken Cray XT5 system at the National Institute for Computational Sciences (NICS) at the University of Tennessee-Knoxville will become the largest memory computational system available for allocation to the NSF computational science and engineering research community. The Kraken computational system is the most powerful TeraGrid resource at 608 teraflops/s with 66,048 cores in 8,256 nodes. As part of the original acquisition under solicitation 05-525, Kraken is being upgraded to over 99K cores and over 1 petaflop/s of peak performance. With this memory upgrade, which increases 3,840 nodes from 8 GB to 16 GB, all 8,256 nodes will have 16 GB of memory for an aggregate of 132 TB of memory. This uniform configuration will allow researchers to efficiently utilize the entire system for a single application and will add flexibility in scheduling all applications on Kraken. For the most scalable applications that require 16 GB/node, the additional memory will effectively double the size of the largest jobs on the highest performance system in the NSF portfolio.This memory upgrade will allow users to run simulations at higher resolution by running at larger scale, increase the fidelity of their simulations by allowing for the incorporation of additional physical processes and/or variables, and allow applications that are limited by their total amount of memory to run on fewer nodes. Many large scale computational science applications are limited by the amount of physical memory available on a node and computational researchers expect to increase their memory requirements as the scale of the computations increase. Increasing the amount of available memory will greatly increase the flexibility in running more advanced multi-scale and multi-component applications, including applications such as multidimensional supernova simulations using CHIMERA, astrophysical and cosmological structure formation with ENZO, nanoscale electronics with OMEN, seismic hazard research with the Anelastic Wave Propagation code (AWP-Olsen), and computational chemistry with NWChem and MADNESS. Broader impacts of this memory enhancement include support for education and outreach via university classes and laboratory workshops at the University of Tennessee-Knoxville, and at Morehouse College and Claflin College. The upgraded system will provide a state-of-the-art platform allowing students to gain experience with multi-core, massively parallel computational techniques.