CEE PhD Defense: Gustavo A. Araújo Rodríguez, "Heterogenous CPU–GPU Parallel Computing for Real-Time Hybrid Simulation"

About

Abstract: Real-time hybrid simulation (RTHS) combines physical experimentation with numerical simulation to investigate the dynamic response of complex engineering systems. However, the fidelity and complexity of the numerical models used in RTHS are constrained by the computational demands of solving the governing equations in real time. Consequently, numerical models in RTHS are often limited in size (e.g., fewer than 1,000 degrees of freedom) and rely on simplified representations of complex physical phenomena, such as soil–structure interaction (SSI). This dissertation investigates heterogeneous CPU–GPU parallel computing to expand the computational capabilities of RTHS within the open-source OpenSees framework. The proposed architecture distributes element and constitutive state determination across CPU processes while accelerating the solution of the assembled system and the numerical integration of the equations of motion on a GPU. Existing GPU-accelerated direct and iterative solvers are integrated through interchangeable backends, together with a Python-delegated interface for external solver libraries. The computational performance of these implementations is evaluated using structural and soil--structure interaction models ranging from thousands to hundreds of thousands of degrees of freedom. To enable computationally efficient time integration, the model-based modified Kolay–Ricles-alpha (MKR-alpha) integration method is reformulated for CPU-parallelized and GPU-accelerated execution. The resulting computational framework is demonstrated through hydrodynamic RTHS of a soil–foundation–bridge system subjected to sequential earthquake and wave loading at the Large Wave Flume facility at Oregon State University. The numerical model explicitly represented nonlinear SSI using continuum finite elements, avoiding the simplified SSI representations often adopted to meet the computational constraints of RTHS. A model with more than 2,000 degrees of freedom satisfied hard real-time constraints, while models containing approximately 4,000–5,500 degrees of freedom achieved soft real-time performance. Beyond RTHS applications, the implemented framework achieved computational speedups in nonlinear dynamic analyses of models containing hundreds of thousands of degrees of freedom. These findings demonstrate the potential of heterogeneous CPU–GPU parallel computing to enable higher-fidelity numerical modeling in RTHS and expand the computational capacity for large-scale nonlinear dynamic analysis.