Collaborative Research: SI2-SSI: Open Source Support for Massively Parallel, Generic Finite Element Methods

点击翻译按钮获取中文摘要

Partial differential equations are used in a wide variety of applications asmathematical models. Their numerical solution is, consequently, of primeimportance for the accurate simulation and optimization of processes in thesciences, engineering, and beyond.The last decade saw the emergence of large and successful libraries thatsupport such applications. While these libraries provide most of what suchcodes need for small-scale computations, many realistic applications yieldproblems of hundreds of millions or billions of unknowns and require clusterswith thousands of processor cores, but there is currently little genericsupport for such problems, limiting access to the many large publiclysupported computing facilities to experts in computational science andexcluding scientists from many fields for whom computational simulation wouldbe a useful tool. This project intends to build the software infrastructure that will allow awide cross section of scientists to utilize these large resources.This project intends to support the software infrastructure for thelarge-scale solution of partial differential equations on massively parallelcomputational resources in a generic way. It will build on two of the mostsuccessful libraries for scientific computing, the finite element librarydeal.II, and Trilinos that provides the parallel linear algebra capabilitiesfor the former. Specifically, we will: (i) Make support for massively parallelcomputations ubiquitous in deal.II; (ii) Research and develop seamless supportfor problems with billions of unknowns in both libraries and improve theinteraction between the two; (iii) Exploit intra-node parallelism on today'sclusters; (iv) Ensure the applicability of our work on a broad basis byimplementing two real-world applications. Both deal.II and Trilinos have large, active and diverse developer and usercommunities, and this project will actively engage these communities throughuser meetings, short courses, regularly taught classes, mailing lists, anddirect contact in focused projects.