Speed Up Weather Prediction on QCT Developer Cloud: A Case Study on Knights Landing Platform

We present the direct performance measurements of two popular weather forecast models, Weather Research and Forecast Model (WRF) and Models for Predictions Across Scales (MPAS) on Intel's Knight Landing Platform (KNL). WRF is widely evaluated over different platforms while the benchmarks of MPAS are still scarce. In this study we measured the running time of WRF and MPAS on the QCT Developer Cloud, both on its KNL-based nodes and Xeon Broadwell-based nodes. We found that for WRF its performance on single KNL node is 1.55 times faster than Broadwell one, while for MPAS is 1.1 times faster. Generally the scalability of two models on a single node is linear, and drops when across multiple nodes. Further optimization might be needed for those two models.