Scalability of Global 0.25° Ocean Simulations Using MOM

• Main Authors: Ward, Marshall, Zhang, Yuanyuan
• Other Authors: National Computational Infrastructure, Fujitsu Australia Limited, Ralf Denzer, Robert M. Argent, Gerald Schimak, Jiří Hřebíček, TC 5, WG 5.11
• Format: Conference Object
• Language: English
• Published: HAL CCSD 2015
• Subjects: ocean modeling; performance profiling; high performance computing; parallel computing; [INFO]Computer Science [cs]; Sea ice
• Online Access: https://hal.inria.fr/hal-01328605; https://hal.inria.fr/hal-01328605/document; https://hal.inria.fr/hal-01328605/file/978-3-319-15994-2_55_Chapter.pdf; https://doi.org/10.1007/978-3-319-15994-2_55

Part 8: High Performance Computing and BigData International audience We investigate the scalability of global 0.25° resolution ocean-sea ice simulations using the Modular Ocean Model (MOM). We focus on two major platforms, hosted at the National Computational Infrastructure (NCI) National Facility: an x86-based PRIMERGY cluster with InfiniBand interconnects, and a SPARC-based FX10 system using the Tofu interconnect. We show that such models produce efficient, scalable results on both platforms up to 960 CPUs. Speeds are notably faster on Raijin when either hyperthreading or fewer cores per node are used. We also show that the ocean submodel scales up to 1920 CPUs with negligible loss of efficiency, but the sea ice and coupler components quickly become inefficient and represent substantial bottlenecks in future scalability. Our results show that both platforms offer sufficient performance for future scientific research, and highlight to the challenges for future scalability and optimization.