A GPU optimization strategy in nonlinear explicit dynamic analysis for reinforced concrete buildings with composite elements

Abstract The explicit integration for nonlinear structural dynamics in finite element analysis (FEA) is inherently decoupled in its algebraic equations, making it well-suited for parallel computation. This paper presents a novel and efficient central processing unit (CPU)/graphics processing unit (GPU) implementation and optimization strategy for the explicit integration of complex tall buildings subjected to seismic loading for the design software YJK. The presence of multiple element types and distinct material constitutive laws in finite element (FE) models of reinforced concrete building structures results in significant computational overhead and branching. In this paper, the calculation-related data for a FE model is reorganized into several data-domains, each corresponding to sole element type and sole material constitutive law. To achieve higher computational performance, a concurrent kernel execution strategy is implemented on the GPU platform. Instead of relying on the default, inefficient kernel scheduler of GPU, we developed an efficient scheduler to maximize GPU utilization. This scheduler first measures resource requirements of each kernel, then ranks and divides them into sub-kernels for concurrent execution. Performance tests on practical engineering project demonstrate that, without compromising accuracy, the proposed optimization strategy achieves up to 328.66 × performance improvement over CPU serial implementation, and up to 4.76 × and 1.59 × improvements over a simpler GPU implementation and the default GPU scheduler, respectively.