基于实验修正的可预测GPU性能评估模型

he great computing power of the GPU makes CUDA-GPU system applied widely in the study of the deological disasters. However, when we use the DEM to simulate the deological disasters with gradual changes, the large-scale computing and excessive output steps make the program runs too long time. Existing performance evaluation model can not predict the running time of the program precisely,whatsmore, in terms of performance optimization, they must valuate the performance of the system at the end of the simulation ,then propose the optimization strategy. This paper presents a new performance evaluation model which is based on the memory access time and processing time.The model use the early running of program as sample to analyze and fix the model,which can help the model to predict the running time of the program precisely.We can also use the model to locate the bottleneck of the program and propose the optimization strategy.