NERSCPowering Scientific Discovery Since 1974

NERSC Initiative for Scientific Exploration (NISE) 2011 Awards

Towards High Performance Urban Canopy Flows

Joe Prusa, Teraflux Corporation

Associated NERSC Project: Continuous Dynamic Grid Adaptation in a Global Atmospheric Model (m612)
Principal Investigator: William Gutowski, Iowa State University

NISE Award: 400,000 Hours
Award Date: March and June 2011

Advanced and flexible modeling capabilities that enable the simulation of polliution transport and dispersion processes possible in complex terrain and urban environments are of interest because the potential effect of chemical and biological agent release on the population within the cities is still a serious concern.

The aim of this proposal is to advance an efficient and scalable approach for pollution transport and dispersion (T&D) within complex urban environments. The urban canopy introduces strong multiscale inhomogeneities spanning spatial scales from sub-meter to kilometers and beyond.

The specific nature of the T&D problem requires relatively high resolution for the simulations to resolve small scale turbulence near the buildings and the effectively long time scale required to capture transport of the pollutants in the full extent of real cities. An Immersed Boundary (IMB) method is used to represent building effects in the model grid. Transport of pollutants is done with either passive tracer or a heavy particle approach which in practice uses a computationally demanding bin model to represent a distribution of particle properties.

In this project we plan to use the multi-scale multi-physics model EULAG which has already been demonstrated to work efficiently upwards of thousands of processors on a variety of high-end architectures. Because urban canopy problems require significant spatial resolution and large domain scales, they are computationally demanding and require extremely efficient use of the largest available resources. Therefore we intend to advance the performance of the current model algorithms in the range of 20K to 100K processor cores, especially paying attention to increasing the peak performance while maintaining high model scalability.