Performance of WRF on IaaS from the public cloud
Public IaaS (Infrastructure-as-a-Service) offers an excellent platform for weather forecasting that rivals in performance to the supercomputers and clusters traditionally used to carry out High Performance Computing (HPC) simulations. By tailoring the deployment and configuration of cloud clusters, IaaS offers greater flexibility in terms of number of cores, fluctuating workloads and long-term storage solutions than traditional clusters. An important financial benefit is the shift from a large upfront investment to capital expenditure. IaaS users also benefit from gaining access to the latest technologies much more quickly than on-premises hardware, which usually require a procurement process. A consequence of these factors is that IaaS allows small organizations to access HPC capabilities traditionally reserved to large businesses, national labs, and research universities.
Among the many weather prediction software tools available nowadays, the Weather Research and Forecasting (WRF) Model is one of the most popular worldwide with over 48,000 registered users from 160 countries. The present evaluation uses the WRF NWSC-3 benchmark specifically developed to assess the performance of the next generation of High-Performance Computing and Storage System. Fig. 1 illustrates the results for the first ring of production jobs, which sits at around 576 cores. Total times results from adding radiation time (⁓25%) plus non-radiation computations (⁓75%). The reference time is from the NCAR (National Center for Atmospheric Research) Cheyenne supercomputer, which as of June of 2020 is ranked number 52nd in the top500 list (www.top500.org).
In addition to performance, any decision to migrate WRF to IaaS from the public cloud must include cost considerations. We strongly recommend a total cost of ownership (TCO) evaluation for organizations interested in this migration as several factors dictate the final cost. As a yardstick, Fig. 2 provides a cost comparison for several clusters from AWS. The cost estimate* only accounts for computational power, and has been normalized based on the ratio between price and performance.
Contact if you have any questions about these results or are interested in the next level of benchmarks (i.e. 1,728 or 3,456 cores). Benchmarks on clusters from different cloud service providers will be available soon.
*Costs are evaluated based on prices in the U.S. East Coast region (Northern Virginia).