The International Conference for High Performance Computing, Networking, Storage and Analysis
Implementation of the Colorado State University Global Cloud Resolving Model on the NSF Blue Waters System.
Authors: Ross P. Heikes (Colorado State University), David A. Randall (Colorado State University), Celal S. Konor (Colorado State University)
Abstract: We have ported key components of the Colorado State University (CSU) Global Cloud Resolving Model (GCRM) to the NSF Blue Waters system. With the model on Blue Waters there are two levels of parallelism to explore. The first is a coarse grain MPI based communication between computational cores. This parallelism works in conjunction with our global domain decomposition. We show scaling characteristics of the Blue Waters system to 40K cores, and include comparisons with other computer systems. The second level of parallelism is a fine scale parallelism which utilizes the NVIDIA Kepler accelerators. This loop based
parallelism directly modifies the numerical operators used within the model using openACC directives. We show that the parallel efficiency of the accelerator strongly depends on the problem size, and have devised modifications to the model to better utilize the accelerators.