Skip to main navigation Skip to search Skip to main content

Porting of an Atmospheric Chemistry Solver to Parallel CPU-GPU Execution

Student thesis: Doctoral thesis

Abstract

Earth System Models (ESMs) are crucial for understanding climate change, providing valuable insights into real-world phenomena and systems, with numerous applications in science and engineering. However, the scientific community requires greater computational power to reduce uncertainties in these models. Significant progress is being made by transitioning to heterogeneous computing systems that combine Central Processing Units (CPUs) and Graphics Processing Units (GPUs). These models can be accelerated notably by optimizing the most time-consuming components, such as chemistry, thereby reducing uncertainties and enhancing our understanding of climate dynamics._x000D_ _x000D_ The primary objective of this thesis is to improve the performance of chemistry solvers in atmospheric models by developing efficient solutions tailored to GPU-based heterogeneous architectures. Specifically, the focus is maximizing GPU performance, enhancing accuracy, ensuring portability, and minimizing development effort by minor modifications to the existing chemistry solver._x000D_ _x000D_ This thesis's critical contribution is developing a CPU-GPU version of the CAMP chemistry solver for the atmospheric model MONARCH, achieving significant acceleration. Furthermore, the developed chemical solver can be easily adapted to solve generic systems of Ordinary Differential Equations (ODEs) on the GPU due to its similarity to the CVODE library used in the CAMP CPU solver._x000D_ _x000D_ The open-source code has been released as an upgrade to CAMP to facilitate portability. The main contributions include a method for seamlessly integrating GPU code into chemistry solvers (Multi-cells), distributing computational loads across GPU threads (Block-cells), and an automatic load-balancing algorithm for concurrent CPU-GPU execution. Notably, the automatic load-balancing algorithm is designed to be simple, ensuring portability while delivering accelerated execution. Furthermore, a GPU linear solver has been developed following the Block-cells approach, enabling further exploration of Block-cells in iterative solvers and facilitating coupling with other systems—an important consideration given the widespread use of linear solvers in scientific and mathematical communities._x000D_ _x000D_ The results demonstrate a 250x speedup compared to the single-threaded CPU version when using four GPUs in a single node. In a node-to-node comparison, utilizing four GPUs versus 80 CPU cores, the speedup achieved is 8.14x. When normalizing the cores used against other state-of-the-art solutions, the speedup is 16x. The substantial speedup achieved with minimal algorithmic changes underscores the significance of the new strategies developed in this thesis.
Date of Award26 Nov 2024
Original languageEnglish
SupervisorMario César Acosta Cobos (Director) & Eduardo Cesar Galobardes (Director)

Cite this

'