•  
  •  
 

Caravel Undergraduate Research Journal

Abstract

Accurately predicting storms and hurricanes is critical to saving lives and reducing economic loss. Therefore, it is necessary to use the most efficient software and hardware technology available in order to improve the performance and fidelity of these predictive mathematical models. For over ten years, the Computational Hydraulics Lab (CHL) at the University of Notre Dame has been involved in developing the high-resolution ADvanced CIRCulation (ADCIRC) storm surge model to predict storm surges in coastal areas. The objective of the work reported here was to port a novel adaption of the parallel ADCIRC code to the state-of-the-art Intel Xeon Phi Co-Processor system (Stampede) at the Texas Advanced Computing Center (TACC) and ideally demonstrate speedup on a set of benchmark calculations. The porting process was accomplished by identifying fine-grained parallelism and vectorizable compute-intensive loops that could be offloaded to the 61-core Xeon Phi co-processors while leveraging their 512-bit wide vector units. Due to transfer latencies, offloading excessive amounts of code can reduce the effectiveness of using the Xeon Phi co-processors so the code was initially profiled in order to identify the code hotspots where the host processors spent the majority of their time. This analysis allowed us to focus the work of OpenMP and Xeon-Phi offloading on the most expensive routines. These routines were modified to take advantage of the full 16 threads available on the host processor, and will ultimately allow us to offload work to the increased number or threads on the partnered Xeon-Phis.

Download

Share

COinS