Investigating Algorithms, Performance, and Scalability in HPC Simulations of Planetary Atmospheric Turbulence on Modern GPUs

Sammanfattning

Increasing the resolution of atmospheric turbulence simulations is a wellknown task for advancing computational geophysics and climate modeling. Efficient execution on modern heterogeneous computing platforms requires development strategies that take into account algorithmic structures, hardware architectures, and parallelization methods. This thesis investigates the portability, performance, and scalability of the NAN atmospheric turbulence simulation on CPU–GPU systems, focusing on the interplay between memory hierarchies, communication topologies, and parallel workloads. Portability on different platforms is achieved by using OpenMP target offloading, letting the simulation run with minimal adjustments. This approach balances codebase compatibility across different hardware architectures with performance, pointing out if and when hardware tuning is really necessary. Strategies on how to treat trade-offs between intra-node and inter-node parallelization are pointed out, presenting how to balance computational throughput, communication cost, and numerical stability. Based on these insights, future performance improvements are proposed, including topologyaware domain decomposition, asynchronous communication, memory optimization techniques, and mixed-precision computation. By performing strong and weak scaling experiments on GPU clusters, it is shown how domain decomposition, stencil-based kernels, and internode communication affect performance. The results indicate that GPU acceleration provides speedups, but the primary bottlenecks are memory bandwidth and communication overhead, particularly at large node counts. These findings provide a framework for understanding the portability, performance, and scaling behavior of large-scale atmospheric simulations and offer practical guidance for optimizing high-performance computing resources, resulting in more efficient and higher-resolution turbulence modeling.