混合并行两步调整遗传策略的软硬件划分算法  

Hybrid parallel genetic strategy with two-step adjustment for HW/SW partitioning

作　　者：侯能 何发智[2] Hou Neng;He Fazhi(College of Computer Science;State Key Laboratory of Software Engineering, Wuhan University, Wuhan 430072, China)

机构地区：[1]武汉大学计算机学院,湖北武汉430072 [2]武汉大学软件工程国家重点实验室,湖北武汉430072

出　　处：《华中科技大学学报（自然科学版）》2017年第12期39-45,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61472289);湖北省自然科学基金资助项目(2015CBF254)

摘　　要：为了提高软硬件划分方法的效率,针对已有遗传算法求解软硬件划分没有结合特定问题处理、不满足约束个体的不足,提出一种混合并行的两步调整遗传算法.采用两步调整策略将不满足约束的个体转换为可行个体,当提高方法的运行效率时,图形处理单元用于计算每个个体的硬件耗费、软件耗费和通信耗费,多核CPU(中央处理器)用于并行执行个体间的调整,流并发传输策略进一步减少CPU和GPU(图形处理器)之间的传输开销.在基准数据集上,与求解该问题的已有方法相比,运行时间和求解质量都有明显优势.实验结果验证了该方法的有效性和合理性.In order to improve the efficiency of hardware/software（HW/SW）partitioning methods and to overcome the shortcoming in which existing genetic algorithm fails to combine with specific problem when dealing with infeasible individuals,a hybrid parallel based genetic strategy with two step adjustment for HW/SW partitioning was presented.Two step adjustment strategy was utilized to transform the infeasible individuals into be feasible.While improving the efficiency of proposed method,graphical processing unit（GPU）was utilized to compute each individual′s hardware cost,software cost and communication cost.Multi-core CPU（central processing unit）was utilized to adjust the infeasible individuals in a parallel way.A stream concurrency strategy was utilized to minimize the transfer overhead between CPU and GPU.Based on the benchmark,when comparing against the existing methods,the proposed method has a significant advantage of solution quality and time.The experimental results indicate the validity and rationality of proposed method.

关 键 词：软硬件划分 遗传算法 两步调整 图形处理单元 多核CPU 

分 类 号：TP302.7[自动化与计算机技术—计算机系统结构]