Making superparameterization flexible for low cloud feedback analysis by focusing turbulence-permitting resolution where it matters most with geographic parallel load balancing

Abstract:

Cloud feedback remains a major challenge in predicting future climate change, especially for low clouds that require horizontal grid spacings of less than 100-m to resolve. Superparameterized simulations are attractive in this context because they strategically undersample space to afford ambitious resolutions that are not approachable in uniform mesh models. However, even SP models cannot yet afford the 3D Large Eddy Simulation (LES) resolution that would be desired to do the most justice to low cloud feedback physics. One obstacle is that SP has historically taken a “one-size-fits-all” approach in that the same CRM grid mesh must be used everywhere, to avoid geographic load imbalance. This is at odds with experience in the cloud-resolving community that different convective regimes merit different computational intensity per unit area. We introduce a way to sidestep the issue and create a geographically flexible superparameterization by intervening in the host climate model’s parallel load balancing. As proof of concept that our scheme can allow reasonable throughput even in situations of extreme geographic work imbalance, we show results from pilot simulations in which one-third of the globe is covered by grid cells using an augmented form of ultraparameterization (UP) while the rest of the grid cells are configured using classical superparameterization (SP). Trade-offs in the optimal work imbalance ratio and communication overhead for performance are discussed. We expect the new load balancing infrastructure can help make SP more generally flexible for scientific inquiry in other areas. For instance, by reducing the cost of classical SP by only superparameterizing the tropics, or making SP three-dimensional where convective momentum transport especially matters, or enlarging the room available for mesoscale storm organization over densely populated areas.