Meso-gamma-scale numerical weather simulations for sub-Saharan Africa via grid-based, distributed computing

Abstract

Numerical simulations at cloud-resolving scales have becoming practical for both research and operational applications due to advances in computing technology. However, deploying such capabilities beyond a limited scale (e.g., extended metropolitan region, large watershed) typically remains out of reach due to the computational cost and the complexity of the systems to support such work. Yet such capabilities are needed to address the local impacts of precipitation events that can impact much broader areas. In particular, convective storms driven by monsoons remain unresolved by current numerical weather prediction systems applied to sub-Saharan Africa. To address this problem, the African Rainfall Project (ARP) was initiated to deploy the community Weather Research Forecast (WRF) model across this region at 1x1 km horizontal resolution on the World Community Grid (WCG). WRF is configured to capture a diversity of geographic conditions in the region with appropriate boundary layer, land surface and cloud microphysics and parameterizations in addition to high vertical and temporal resolution. WCG provides a fully distributed computational environment that crowd-sources unused computing power from volunteers’ devices and donates it to scientific projects. As such, all computations must be embarrassingly parallel, which creates a challenge for models like WRF. Hence, each instance of WRF must operate serially on a volunteer’s device. To address the regional-scale simulations, sub-Saharan Africa is decomposed into individual 52 by 52 km domains at 1x1km as the third nest in two-way telescoping grids with common centroids. The outer domains are at 3 and 9 km resolution, respectively with the same vertical resolution. Each 48-hour simulation is done as a cold-start forced by reanalysis with output saved every 15 minutes. The collection of these simulations will cover at least one year to capture seasonal variations. Since there is no operational imperative, the ability of typical volunteer’s system to compute each simulation in several hours is practical. Scaling is achieved with many thousands of systems being deployed simultaneously. With this decomposition, over 35000 overlapping domains cover the region. During post-processing, the individual simulations are stitched together to create a consistent, single output for over for the period of study. Although the focus is precipitation, the simulations provide additional standard output for 2m temperature and 10m horizontal wind velocity, for example. We will report on the results to date and validation in comparison to in situ (e.g., from TAHMO, www.tahmo.org) and remotely sensed observations as well as conventional WRF deployments for a large computational domain covering a small subset of the region.