Nonlinear Earthquake Response Analysis Of High-rise Reinforced Concrete Structures Based On Heterogeneous Platform

With the trend of refined modeling of high-rise building structures, millions of degrees of freedom(DOF) are used in the analysis of the spatial structural system, which leads to intensive computation, so an efficient, high-performance of simulation platform is needed. However most simulation platforms are based on CPU serial pattern which was implemented on computers, large time consuming becomes a bottleneck of dynamic structural analysis. How to establish high-performance low-price simulation platform for the finite elem ent analysis(FEA) becomes the urgent problem which is needed to address. The macro-model used for nonlinear FEA can be adapted to reduce the computational intensity and time efficiently. However, these models may not adequately capture all nonlinearities of spatial structure and sometimes the results of the analysis are not accurate. In order to achieve a higher computing accuracy and save calculation time in the process of elastic-plastic time history analysis, a heterogeneous simulation platform combination of CPU and GPU was developed. GPU parallel algorithm of nonlinear FEA program and software were developed. The main contents of the dissertation are as follow:(a) The development and validation of CPU-GPU heterogeneous platform. Based on CPU and GPU hybrid programming, CPU is used to deal with sequential data operations and logic control, while GPU is used to deal with large-scale data processing. Similar to the concept of parallel divide-and-conquer scheme, the basis of parallel implementation is used to establish the CPU-GPU heterogeneous platform. Based on CUDATM programming models, the GPU-based parallel version of conjugate gradient algorithm(CG) and preconditioning conjugate gradient algorithm(PCG) solver are develolped, respectively. These solvers could effectively solve the large sparse system of equations. Several optimization strategies are employed to further improve the computing efficiency of CPU-GPU heterogeneous platform. The banchmark tests demonstrated the reliability and effectiveness of heterogeneous platform.(b) Study on fiber model of nonlinear analysis method based on CPU-GPU heterogeneous platform. The fiber model is improved to address the interaction between axial force, bidirectional shear, biaxial bending, and torsion. The damage evolution of concrete specimens confined by cuboid stirrups is presented experimentally. Then a damage evolution equation and plastic damage constitutive model of stirrup-confined concrete are proposed. A GPU-based incremental displacement algorithm is introduced to deal with negative stiffness problems. In the view of finite element discretization, the thread-level parallelism was carried out by mapping the data onto a Stream Processor as a thread. Combining with the fiber model, material model and displacement control algorithm, a GPU-based parallel FEA program is developed to speedup the fiber beam-column element for nonlinear analysis of reinforced concrete structures.(c) Study on layered shell model of nonlinear analysis method based on CPU-GPU heterogeneous platform. The FE matrices of shell elements are obtained by combiantion of plane stress elements and Mindlin plate elements. Biaxial hysteresis constitutive model of concrete which is suitable for simulating the shear-walls and the slabs are established. A GPU-based Newton-Raphson iteration algorithm is proposed to realize parallel solving nonlinear system of equations on CPU-GPU heterogeneous platform. The shell element model can reflect nonlinear complex spatial behavior of the shear-walls. The results of shell element models show a good agreement with the experimental data.(d) Study on elastic-plastic time history analysis of high-rise reinforced concrete structures based on CPU-GPU heterogeneous platform. A spatial model of the high-rise reinforced concrete structures subjected to earthquake loading is proposed. In this spatial model, the beams and columns are simulated by the fiber model, while the shear-walls and slabs are simulated by layered shell model, also uniaxial or biaxial materials nonlinearity is considered. Parallelization strategies for implementations of Newmark-beta algorithm are presented and an entire GPU-based parallel elastic-plastic seismic response analysis program is developed, and implemented within a highly memory-efficient element-by-element(EBE) technique which performs “global” iterative solution of the matrix-vector operation at the “element” level. The validity of the program is verified by comparing with results of shaking table test. Applications of CPU-GPU heterogeneous platform to elastic-plastic time history analysis of two high-rise RC frame tube structures demonstrated that the proposed models and algorithms take full advantage of the parallel GPU technologies. And such high accuracy, high efficiency, good portability and strong adaptability simulation platform shows a significant prospect.