针对脑网络研究中需计算的网络数目过多造成执行时间过长的问题,提出一种基于多核CPU(central processing unit)的并行计算方法.通过SPMD(single program multiple data)机制利用CPU的多核同时执行,实现并行计算多个网络的属性,利用循环打包方法降低SPMD机制中循环控制的时间,得到并行计算多个网络指标的时间,同串行计算时间相比,可得到此方法的并行计算效果.与传统并行单个算法的策略相比,该方法利用不同脑网络之间计算的独立性,采取同时计算多个网络的策略.在一台多核CPU的主机上,分析CPU核数、网络节点规模这两个因素对网络指标计算并行效果影响.在利用12个CPU核并行计算网络节点规模为3000的指标时,加速比均达到2以上,其中效果最好的是网络同配系数的计算,加速比达到6倍以上.实验结果表明,基于SPMD机制和循环打包方法的并行计算架构对脑网络指标计算的并行效果显著,加速比随着CPU核数、网络节点规模的增长呈上升趋势.%Since the time consumption of computing the index of brain network is excessively high caused by its number,a paral-lel computing method based on multi-core CPU was proposed.The same index of multiple networks was computed in parallel through the SPMD mechanism using the multi cores of CPU.A kind of loop packing method was established to reduce the time of cycle control in SPMD.The parallel time and the serial time were recorded,and the parallel computing speedup was obtained. Compared with the traditional method of parallel algorithm,this method calculated the multiple networks at the same time since it taking advantages of the independence of calculating the index between different brain networks.In a multi core CPU host,the effects of the CPU kernel number and the network node scale on the parallel performance were analyzed.Using 12 CPU cores to parallel compute the network of 3000 nodes,the speedup reached 2,and the best effects were more than 6 times for the calcula-tion of the assortativity coefficient.Experimental results show that the parallel computing architecture based on SPMD mecha-nism and the method of packaging loop body has significant effects on the performance of the brain network.And the number of CPU cores and the scale of network nodes are positively correlated with the speedup of parallelism separately.
展开▼
