The Weather Research and Forecasting (WRF) Model is a key tool for atmospheric research and short-term forecasting with high computational demands. This study investigates the parallel execution of WRF in a high-performance computing (HPC) environment, focusing on the balance between MPI ranks and OpenMP threads along with the role of (Non-Uniform Memory Access (NUMA) affinity. We evaluate how different workload distributions affect computational efficiency by simulating weather patterns over Colombia with varying domain sizes. Our findings highlight the significant impact of process placement and thread configuration on reducing simulation times, offering practical guidelines for enhancing WRF performance in HPC environments. Incorrect placement can lead to suboptimal memory access patterns and unnecessary CPU context switching, all of which degrade performance. This study demonstrates that better performance can be achieved by balancing computation and communication times.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Parallel Computing Strategies in WRF: The Role of MPI, OpenMP, & NUMA Affinity

  • Diego A. Roa Perdomo,
  • Esteban Hernandez,
  • Xiaoming Li,
  • Kevin A. Brown

摘要

The Weather Research and Forecasting (WRF) Model is a key tool for atmospheric research and short-term forecasting with high computational demands. This study investigates the parallel execution of WRF in a high-performance computing (HPC) environment, focusing on the balance between MPI ranks and OpenMP threads along with the role of (Non-Uniform Memory Access (NUMA) affinity. We evaluate how different workload distributions affect computational efficiency by simulating weather patterns over Colombia with varying domain sizes. Our findings highlight the significant impact of process placement and thread configuration on reducing simulation times, offering practical guidelines for enhancing WRF performance in HPC environments. Incorrect placement can lead to suboptimal memory access patterns and unnecessary CPU context switching, all of which degrade performance. This study demonstrates that better performance can be achieved by balancing computation and communication times.