Georgia Institute of Technology
Accelerating Legacy Computational Fluid Dynamics (CFD) Solvers for Modern GPU Hardware
Abstract
dc:description.abstractLegacy combustion solvers are often memory bound by nature; these solvers are written for CPU cores, often faces performance bottlenecks when porting to GPUs because the control flow in CPUs is completely different to GPUs, as CPUs are particularly Latency-focused, while GPUs are throughput-focused. With the rise of AI and ExaScale computing clusters all around the world, Graphic Processing Unit (GPUs) have almost become the Powerhouse of modern computational research in both academia and industries because of their high-bandwidth utilization, offering higher speedups and massive parallelism with finer controls. This work presents the study of GPU acceleration and optimizations of the LESLIE, a finite volume based multi-block multi-physics compressible Navier-Stokes structured solver used for Large Eddy Simulations with Linear Eddy Model. The work investigates the transition to OpenACC directive-based GPU offloading, examining critical parameters for GPU selection, kernel-level optimization techniques, profiling studies on occupancy versus register pressure, bandwidth utilization, and memory coalescence for complex physics kernels. A strong scaling analysis is presented for fixed grid geometries, along with a GPU-native algorithm for data packing and unpacking through halo exchanges, and deeper analysis on warp-level utilization. The overall GPU codebase is studied on three different cases a periodic box, temporal mixing layer, and solid fuel ramjet. Observed with an end-to-end performance gain of around 4X on solid fuel ramjet configurations on 4 NVIDIA H200 GPUs compared to 1113 Intel CPU cores, and around 9X on temporal mixing layer with 11 species chemistry. The study concludes being FP64 compute and purely bandwidth bound solver, rewriting the subroutines to match the GPU architecture will improve in controlling the memory traffic and results in end-to-end performance gains.
Degree
thesis:*- Name thesis:degree_name
- Computational Science and Engineering, MS
- Department thesis:degree_department
- Mechanical Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kodamalla, Vijay Prabhas
- Advisors dc:contributor.advisor
-
- Jain, Suhas
- Menon, Suresh
- Committee member dc:contributor.committeemember
-
- Oefelein, Joseph
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81650