基于GPU并行的功能脑网络属性分析方法

Parallel analysis methods of functional brain network characteristics based on GPU

为实现大规模功能脑网络拓扑属性的高效计算,提出基于GPU并行的脑网络属性分析方法。采用统一计算设备CUDA架构,属性分析方法中的计算密集型操作由GPU完成。以功能脑网络为对象,在GPU型号为NVIDIA Quadro K4200的工作站上对该并行方法进行模拟,将该方法与基于单程序多数据SPMD机制的脑网络属性分析方法进行比较。实验结果表明,当网络节点数大于1000时,该方法具有更高的计算性能。

To achieve the efficient computation of the large-scale functional brain network topological characteristics,GPU-based parallel methods of brain network characteristics were proposed.Based on the compute unified device architecture,the computation intensive operations in network characteristics methods were performed by GPU.To analyze the parallel performance of the proposed methods,experiments were carried out on a workstation with a NVIDIA Quadro K4200 GPU.Experiments demonstrate that the proposed methods improve the computation efficiency of functional brain network characteristics.Compared with the SPMD-based parallel methods of brain network characteristics,results indicate that GPU-based parallel methods have better parallel performance when the number of functional brain network nodes is more than 1000.