TOWARDS ANALYSIS-DRIVEN MULTIPHYSICS SOFTWARE ARCHITECTURE
TITLE:
TOWARDS ANALYSIS-DRIVEN MULTIPHYSICS SOFTWARE ARCHITECTURE
DATE:
Friday, April 10th, 2009
TIME:
3:30 PM
LOCATION:
GMCS 214
SPEAKER:
Damian Rouson, Ph.D., P.E., Principal Member of Technical Staff, Scalable Computing R and D, Sandia National Laboratories
ABSTRACT:
Leading-edge computational science increasingly requires multidisci- plinary teams constructing multiphysics models. As the number of codes in a coupled solver grows, the process of scaling up to hundreds of pro- gramming units (e.g., procedures, classes, components, or developers) impacts pro ject budgets and timelines at least as much as the process of scaling up to hundreds of execution units (e.g., processors, processes, threads, or cores). Put simply, distributed development matters as much as distributed computing. After making this argument quantitative by combining Amdahl’s Law with the Pareto Principle, this talk will bring an assortment of analytical tools to bear on the question, “How does the high-level organization of a code impact development-time processes?” The tools range from established ob ject-oriented design metrics to new computational complexity estimates applied to debugging and informa- tion entropy arguments applied to developer communications. The ap- proach to be analyzed involves defining an abstract data type (ADT) calculus comprising scalar-, vector- and tensor-field ADTs that support a set of user-defined algebraic and integro-differential operators. It will be demonstrated that ADT calculus leads to highly cohesive, loosely cou- pled abstractions that reduce bug search times and interface information content. The talk will also present the application of ADT calculus to problems ranging from particle dispersion in magnetohydrodynamics to quantum vortex interactions with normal fluid turbulence and electro- magnetic wave scattering in the atmospheric boundary layer.
HOST:
Jose Castillo
DOWNLOAD: