计算机工程与应用
計算機工程與應用
계산궤공정여응용
Computer Engineering and Applications
2015年
22期
70-76
,共7页
图形处理器(GPU)%遗传算法%自适应邻域%计算统一设备架构(CUDA)%Guassion分布
圖形處理器(GPU)%遺傳算法%自適應鄰域%計算統一設備架構(CUDA)%Guassion分佈
도형처리기(GPU)%유전산법%자괄응린역%계산통일설비가구(CUDA)%Guassion분포
Graphic Processing Unit(GPU)%genetic algorithm%adaptive neighborhood%Compute Unified Device Archi-tecture(CUDA)%Gaussian distribution
提出了三种新的GPU并行的自适应邻域模拟退火算法,分别是GPU并行的遗传-模拟退火算法,多条马尔可夫链并行的退火算法,基于BLOCK分块的GPU并行模拟退火算法,并通过对GPU端的程序采取合并内存访问,避免bank冲突,归约法等方式进一步提升了性能。实验中选取了11个典型的基准函数,实验结果证明这三种GPU并行退火算法比nonu-SA算法具有更好的精度和更快的收敛速度。
提齣瞭三種新的GPU併行的自適應鄰域模擬退火算法,分彆是GPU併行的遺傳-模擬退火算法,多條馬爾可伕鏈併行的退火算法,基于BLOCK分塊的GPU併行模擬退火算法,併通過對GPU耑的程序採取閤併內存訪問,避免bank遲突,歸約法等方式進一步提升瞭性能。實驗中選取瞭11箇典型的基準函數,實驗結果證明這三種GPU併行退火算法比nonu-SA算法具有更好的精度和更快的收斂速度。
제출료삼충신적GPU병행적자괄응린역모의퇴화산법,분별시GPU병행적유전-모의퇴화산법,다조마이가부련병행적퇴화산법,기우BLOCK분괴적GPU병행모의퇴화산법,병통과대GPU단적정서채취합병내존방문,피면bank충돌,귀약법등방식진일보제승료성능。실험중선취료11개전형적기준함수,실험결과증명저삼충GPU병행퇴화산법비nonu-SA산법구유경호적정도화경쾌적수렴속도。
Three new GPU-based parallel simulated annealing algorithms with adaptive neighborhood are proposed in this paper. They are parallel genetic-simulated annealing algorithm based on GPU, parallel annealing algorithm with multiple Markov chains, and parallel annealing algorithm based on block. Several novel strategies adopted in these algorithms such as coalescent memory access, avoiding bank conflict, and reduction improve the performance. The experiments tested on 11 typical benchmark functions show the new three algorithms have better accuracy and faster convergence speed than the nonu-SA algorithm.