摘要
本文研究LogP模型上一类蝶式计算中的通信问题.以FFT的并行计算为例,通过仔细安排消息的发送顺序,使得由有限带宽引起的延迟与局部计算重叠,在g-logg+1≤logp(p为处理器数,g为带宽因子)的条件下,只要输入长度n满足最基本的要求(n≥2p2),g便被完全隐含于局部计算中,算法时间复杂度可达到最优.最后与文献[1]的结果比较,分析了它们的优缺点及各自的适用范围.
This paper discusses communication problem of butterfly computationusing the LogP model of computation. In order to mask communication time causedby bandwidth-limited as much as possible,the authors carefully overlap interprocessor communication steps with local computation, and give FFT algorithm as an example. Analysis shows that the technique is suitable for butterfly computation. Letn22p' and g-logg+1≤logp, the running time of the algorithm is optimal. Finally, the author's result is compared with Sahay's and their adaptabilities are analyzed respectively.
出处
《计算机学报》
EI
CSCD
北大核心
1997年第8期695-701,共7页
Chinese Journal of Computers
基金
国家教委博士点基金
关键词
蝶式计算
LOGP模型
通信
算法
Butterfly computation, FFT permutation, in-driven.