Computational Considerations for the Community-Based Shallow Water Model, ADCIRC, on HPC Clusters Utilizing the Intel® Xeon Phi Processor

The University of Oklahoma was recently selected to host an Intel® Parallel Computing Center, with the objective of improving the performance of the community-based shallow water model called ADCIRC (a 2D/3D coastal circulation and storm surge model whose details are given at http://adcirc.org) on HPC clusters utilizing the Intel® Xeon Phi processors. Recent developments for ADCIRC have focused on its usage in real-world applications. Examples of these applications include the following: (1) After Hurricane Katrina, Congress directed FEMA to use ADCIRC to develop their new flood inundation maps for US coastal areas; (2) After the Fukushima disaster in Japan, the Nuclear Regulatory Commission is revising guidelines to use ADCIRC in flood studies for nuclear power plants; (3) NOAA, DHS, the Corps of Engineers and others use ADCIRC for prediction of storm surge and flooding; (4) The State of Louisiana uses ADCIRC as a planning tool for their coastal restoration plan; (5) The Corps of Engineers, an original ADCIRC sponsor and continued supporter, uses ADCIRC in many applications, including design enhancements to New Orleans' hurricane protection system and updating a tidal database for the continental US; (6) After the Deepwater Horizon oil spill, a highly efficient particle tracking routine was coupled to the ADCIRC hydrodynamics to predict surface migration of the resulting oil slick; (7) A software superstructure, called the ADCIRC Surge Guidance System (ASGS), was developed for real-time applications by ingesting meteorological and hydrologic data and producing real-time predictions of flooding extents due to tropical and extra-tropical storms (http://nc-cera.renci.org). In order to produce these high-resolution, real-time simulations, ADCIRC must be computationally efficient. For the current MPI version on parallel HPC architectures, ADCIRC utilizes the METIS software to partition the global problem into sub-domains, one per CPU core. Scaling studies show that ADCIRC is very efficient. OU's Intel® Parallel Computing Center focuses on further ADCIRC modifications via OpenMP threading and vectorization, using Xeon Phi natively as a host, as a mechanism for optimizing the code for both Xeon Phi and Xeon CPUs. In this presentation, results from a code performance analysis and subsequent code optimization will be shown.