J41.1 Using New Programming Models to Improve the Scalability and Performance of Near-Real-Time High-Resolution Coastal Modeling

Wednesday, 10 January 2018: 10:30 AM
Room 12B (ACC) (Austin, Texas)
Hartmut Kaiser, Louisiana State Univ., Baton Rouge, LA

Coastal modelling in general and Hurricane storm surge forecasting in particluar face the issue of having to bring existing codes to modern computing architectures. Those codes have been often developed by domain scientists over a long period of time. The ADCIRC storm surge forecasting system which is widely used in the community is a FORTRAN code which has gained wide acceptance. ADCIRC addresses the effects and interactions of processes such as winds, tides, waves and currents on hydrodynamic interfaces using the two-dimensional (vertically-integrated) and three-dimensional versions of the Shallow Water Equations, widely used in coastal modeling.
The STORM project brings together a group of researchers from computer science, coastal science and coastal engineering to create a sustained effort on software development activities that broaden ADCIRC from a somewhat static model based on a single, aging solution algorithm and parallelization strategy into a dynamic, multi-algorithmic modeling environment that takes advantage of recent transformational advances in parallel computing.
Modern computers, such as the upcoming machines based on the the latest Xeon-Phi (KNL) architecture pose a completely new set of computational challenges, such as handling deep memory hierarchies with special high bandwidth memory blocks, and hundreds of processing units per node. All of the existing programming models still have to be shown to be efficiently usable for those.
The STORM project integrates the existing ADCIRC code base with a new task-based runtime system, which is based on fine-grain parallelism, a global address space, and a very asynchronous programming model. This runtime system, HPX has shown to efficiently utilize large amounts of parallelism with scaling results beyond existing practice.
This talk presents the first results from the integration and development work in the context of the STORM project.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner