GPU acceleration of an exascale-ready parallel Fourier-based Poisson solver

Details

Supervisors

Chatelain, Philippe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil mécanicien, à finalité spécialisée

Abstract

enHigh-performance computing (HPC) applications increasingly rely on efficient GPU-to-GPU communication. In distributed Fast Fourier transform (FFT) solvers such as FLUPS, communication between GPUs often becomes the bottleneck. This effect is particularly pronounced for small grid sizes, where the limited message sizes prevent the network from reaching peak bandwidth and fixed costs such as synchronization and latency dominate. For large grid sizes, the bulk-synchronous communication strategy is effective, as the larger messages amortize synchronization costs. However, for smaller problems this strategy incurs excessive overhead, leading to reduced effective performance despite high theoretical network bandwidth. This thesis explores and implements multiple strategies to improve inter-node GPU communication in FLUPS. A key focus is the development of a finer-grained communication pattern, effectively reducing the number of host side synchronizations. This is achieved by improving the potential overlap of communication and computation, reducing the generated overhead and enabling more effective bandwidth saturation. Profiling with Nvidia Nsight Systems was used to quantify the performance impact and to evaluate the potential benefits of GPU-triggered communication over the host-triggered approach. The study also evaluates trade-offs between implementation complexity and achievable performance improvements, considering both MPI-based and native Libfabric solutions. Overall, this study provides a quantitative assessment of potential performance gains and proposes practical strategies to realize them in real-world HPC applications.