Study On CUDA-based Parallel Algorithm Of The Dynamic Load Identification

When we process dynamic load identification of the large and complex structure, the solutionprocess is very time-consuming. Thus, it is difficult to obtain satisfactory results in traditional serialalgorithm. This paper presented the CUDA-based parallel algorithm of the dynamic loadidentification. Because of the multiprocessor of GPU, when we deal with large amounts of data withthe method of parallel algorithm, we can improve the efficiency of the solving evidently. In order toimprove the solution speed, this paper made the following research:First, this paper described the theory of generalized chebyshev orthogonal polynomials, andderived the identification formula of frequency domain identification dynamic load distribution andtime-domain identification dynamic load distribution by the way of function fitting, specially thispaper focused on the identification formula of beam and sheet.CUDA is a kind of new parallel computing architecture software and hardware based on GPU.Threads are used to control the implement of the program. Because of the good parallelism of finiteelement method, parallel computing technology CUDA-based can be well applied on finite elementanalysis. This paper focused on the establishing of CUDA-based parallel computing platform, theprogramming model of CUDA, the thread structure of CUDA, the way of optimize CUDA code andthe standard to estimate the quality of the CUDA code.Finally, based on the above studies, this paper programmed the module of frequency domainidentification dynamic load distribution and time-domain identification dynamic load distributionwith C/C++. The simulation example proved the accuracy and speedup of parallel algorithms. Thispaper also studied on the location identification of concentrated loads, based on the number of loadswere known. This paper programmed the parallel code of location identification with he method ofdomain decomposition.