Title: Taming Complexity in Multiphysics Software Design Speaker: James C. Sutherland, University of Utah Date/Time: Monday, July 27, 2009 at 2:00pm Location: CSRI Building, Room 90 (Sandia NM) Brief Abstract: Multiphysics simulation software is plagued by complexity stemming from nonlinear coupling. In addition to nonlinear coupling, such software often must support many models, each of which may require a different set of transport equations, constitutive models, and equations of state. Strong coupling, together with a multiplicity of models, leads to complex algorithms and rigid software. The rigid, complex software is due to design that focuses on algorithms. In this talk, we will discuss an alternative programming paradigm where programmers focus not on the algorithm, but on the data. Mathematical expressions are reflected in software in a way that directly exposes data dependencies, and graph theory is employed to automatically generate an algorithm. This approach allows problems with very complex dependencies to become entirely tractable, and removes virtually all logic from the algorithm itself. Changes are highly localized, allowing model developers to implement code without requiring a detailed understanding of any algorithms. It provides a natural framework to achieve model adaptivity, where one model suite may be substituted for another dynamically; adding or removing transport equations, modifying constitutive models, etc. and automatically regenerating a new algorithm. Furthermore, this approach enables efficient algorithmic parallelization via threads. By exposing dependencies in the algorithm explicitly, thread-based parallelism can be implemented through algorithm decomposition, thus providing a basis for exploiting parallelism that is independent from and complimentary to domain decomposition approaches. We will present recent scaling results and discuss how this approach compliments existing MPI-based frameworks such as SIERRA.CSRI POC: Roger Pawlowski, (505) 284-3740 |