A VFSA Scheduler for Radiative Transfer Data in Climate Models

Scheduling and load-balancing techniques play an integral role in reducing the overall execution time of scientific applications on clustered multi-node systems. The increasing computational complexity and higher modeling resolutions of scientific applications makes balanced resource allocation a high priority. While many existing load-balancing approaches attempt to reduce cluster communication overhead by statically estimating job task allocation to the system configuration, there exist significant opportunities to integrate domain-specific information into adaptive scheduling strategies. One such domain is that of atmospheric geophysical data. The evolution of Atmospheric General Circulation Models (AGCMs) and the atmospheric geophysical data on which they rely deem it necessary to revisit prior work in the area of load-balancing. We present a technique that employs Very Fast Simulated Annealing (VFSA) to reduce the load-imbalances associated with radiative-transfer processes found in AGCMs. The technique is evaluated within the Load Balancing and Scheduling Framework (LBSF), a simplification of a climate model that isolates the computations of climate models that relate to load-balancing. Results of the VFSA technique are encouraging when compared to existing algorithms in a distributed AGCM.