Encoded: In-Depth Analysis of MPI/OpenMP Parallelization Tuning for AWS HPC WRF Simulations

• Performance does not improve as MPI ranks are subdivided, with a maximum execution time difference of approximately 22% between the slowest configuration (2 nodes 64:3) and the fastest configuration (3 nodes 24:8)
• The optimal parallelization range was identified as 24:8 to 48:4, and high CPU utilization does not necessarily indicate high efficiency
• In pure spin mode, overhead caused by CPU occupancy during communication waiting is the main cause of performance degradation

Experimental Environment and Constraints

• Used hpc7g, hpc6a, and hpc8a instances in a WRF v4.4.2, 1,000m resolution, 10-minute interval simulation environment
• The upper limit of MPI processes was determined to be 180 based on the number of grid cells (18,240 cells) in the parent domain d01
• Performance was measured with Hyper-Threading OFF, based on 1 vCPU = 1 physical core

Performance and Communication Analysis

• The 2-node 96:2 configuration showed an 84.6% increase in transmitted packets compared to 24:8, causing latency due to decreased average packet size
• Impact of communication processing overhead and accumulated latency resulting from increased small packets, rather than bandwidth saturation
• Variation was very small upon repeated runs (maximum 0.10 minutes for hpc8a 3 nodes), making differences between configurations clear

AWS HPC Utilization and Limitations

• Access to the latest hardware and up to 300 Gbps EFA is possible without initial on-premises investment
• HPC-specific instances cannot be used in the Seoul region, and data governance review is required when using overseas regions
• Capacity Reservations are recommended to secure nodes during peak times for new instance demand