Viability of Cloud-Computing for Real-Time Numerical Weather Prediction over Regional Domains

The use of cloud-computing services, like those provided by Google, Amazon, and Microsoft, is growing rapidly within the atmospheric sciences, including for numerical weather prediction (NWP). However, the transition from using on-premise high performance computing (HPC) systems to cloud services can be a challenge. All clouds are not equal and cost structures, computing options that are available, and underlying network infrastructure vary between providers. We design a virtual-HPC cluster on the Google Cloud Platform and test the performance of the Weather Research and Forecasting (WRF) model to find configurations of the cloud environment that allow us to meet two main requirements of real-time NWP: (1) fast forecast completion (timeliness), and (2) economic cost-effectiveness when compared with traditional on-premise HPC hardware. We find that the Google Cloud Platform is viable for our real-time, regional NWP needs by using three specific configurations: (1) WRF is compiled with the Intel compiler collection, (2) no more than 8 virtual CPUs are used per virtual machine, and (3) virtual machines should use the lowest memory option available. Additionally, using cloud services for real-time NWP has the added benefits of increased hardware reliability and easy expandability.