Incored: AWS HPC Instance WRF Benchmark: hpc7g Ranks First in Cost Efficiency

Renewable energy company Incored released results comparing the performance and cost of hpc7g, hpc6a, and hpc8a while running WRF weather simulations on AWS HPC dedicated instances. The key conclusion is that instance selection should be based on total cost to complete within a target time, rather than hourly unit price.

Execution Time and Cost Comparison

Experimental results showed that the hpc7g 10-node configuration ranked first with an execution time of 90.99 minutes and a cost of $25.53. The hpc8a 3-node configuration recorded 97.25 minutes and $38.51, placing second, while the hpc6a 15-node configuration was the most expensive and slowest at 103.83 minutes and $74.75. Although hpc8a reduced costs to roughly half of hpc6a by cutting the node count from 15 to 3, it did not surpass the cost efficiency of hpc7g.

MPI/OpenMP Optimization and Limitations

In hpc8a tests, performance varied depending on how the 192 cores per node were distributed between MPI processes and OpenMP threads. For the 3-node setup, the 24:8 ratio was the fastest, with a maximum execution time difference of 21% observed across all configurations. Due to WRF's tightly-coupled MPI workload characteristics, performance does not scale linearly with unlimited core increases, and excessive node deployment was inefficient due to the domain decomposition limit (180 processes).