Numerical weather prediction on AWS Graviton2
the Weather research and forecasts (WRF) is a Numerical Weather Prediction (NWP) system designed to meet the needs of both atmospheric research and operational forecasting. The WRF model serves a wide variety of meteorological applications at scales ranging from meters to thousands of kilometers. WRF is one of the most widely used NWP models in academia and industry with over 48,000 registered users in over 160 countries.
With the release of Arm-based AWS Graviton2 Amazon Elastic Compute Cloud (EC2) Instances A common question has been how these instances perform on large scale NWP workloads. In this blog, we’ll present the results of a standard WRF benchmark simulation and compare three different instance types.
AWS Graviton2 The processors are custom designed by AWS using 64-bit Arm Neoverse cores to provide excellent value for money for cloud workloads running in Amazon EC2. These instances are powered by 64 physical-core AWS Graviton2 processors that use 64-bit Arm Neoverse N1 cores and custom AWS-engineered silicon built using advanced 7-nanometer manufacturing technology.
As NWP models often benefit from a high-speed network, we will be evaluating the C6gn.16xlarge (64-core Graviton2 instance) and C5n.18xlarge (36-core Intel Skylake-based instance). These two instances have 100 Gbps network bandwidth and support Elastic fabric adapter (AGE). To identify the performance achieved through the increased networking capabilities of C6gn, we are also evaluating C6g. The C6g instance has the same characteristics as the C6gn instance apart from the increased network capacities.
The benchmark case used for this blog is UCAR CONUS 2.5 km dataset for WRFv4. We use the first 3 hours of this 6 hour simulation, a 2.5 km resolution case covering the continental United States domain (CONUS) from November 2019 with a 15 second time step and a total of d ‘approximately 90 million grid points. Note that in the past WRFv3 has generally been compared using a similar 2.5 km CONUS dataset, however, WRFv3 models are not compatible with WRFv4. Despite the similar name, this is a different model of this dataset.
To learn more about WRF benchmarking on Graviton2, read the full blog here.