Interactive Weather Simulation and Visualization on a Display Wall with Many-Core Compute Nodes

摘要

Numerical Weather Prediction models (NWP) used for operational weather forecasting are typically run at predetermined times at a predetermined resolution and a fixed geographical region. The period between each run is a function of waiting for observational data and the availability of compute resources. The resolution is a function of the geographical region, the available processing power and operational forecasting time constraints. The geographical region is defined by being a region with known need or interest for forecasts. These characteristics make it hard to interactively produce and visualize on-demand high-resolution forecasts for a small and arbitrarily located region. This paper documents a system achieving this, using a high-resolution tiled 22 mega pixel display wall, a 16 node PC cluster and a HP BL 460c blade server with two quad core processors. We document the performance characteristics experimentally. The results show that using 10 km resolution background data, the system produces a 6 hour forecast for a 117 x 123 km small region with 3 km resolution, in 3 minutes. Visualizing the forecast takes between 3-75 seconds. An informal survey among operational forecasters indicate that the majority is willing to wait up to 3 minutes for higher resolution forecasts. This paper identifies and documents some of the bottlenecks and computational challenges created by combining interactivity and traditional batch oriented computing. The main bottlenecks in the system are identified as the execution time of the NWP and the preparation of data for visualization.