Michalakes Envisioned Thesis Abstract

GPU Acceleration of Numerical Weather Prediction

Continued scaling of numerical weather prediction performance is stalling as the historic doubling of microprocessor performance begins to flatten out for conventional architectures. Graphics Processing Units and the Cell Broadband Engine accelerators show great potential for high-performance at low cost and low power, provided NWP models can be adapted to these architectures. Key concerns are the need to reprogram these models and how to reprogram them efficiently while maintaining usability, extensibility, and maintainability. This thesis presents a methodology for adapting weather and climate applications to co-processor accelerators such as GPUs, the Cell processor, and FPGAs; develops domain-specific programming abstractions that facilitate implementing weather and climate applications for GPU acceleration in performance portable fashion; employs the methodology and programming abstractions to implement key computational kernels within actual weather and climate codes; and benchmarks and evaluates quantitatively the performance gains and associated overheads associated with using co-processor accelerators.