分布式系统中任务分配的一种结点自适应算法-fansmale.PDF

26 　3 　　　　　　　　 Vol.26 No.3 2003 3 CHINESE JOURNAL OF COMPUTERS Mar.2003 　 闵　帆　石　兵　杨国纬　周明天 (　 610051) 　　: , . , , , ., . 　;;;; TP301 A Self-Learning Algorithm forIPUs in Distributed Systems MIN Fa 　SHI Bi g 　YANG Guo-Wei 　ZHOU Mi g-Tia (Comp uter Science Academy , University of E lectronic Science and Technology of China, Chengdu 610051) Abstract　I a distributed system with ma y i depe de t processi g u its (IPUs), such as a computer et- work, o e of the most importa t thi gs is to establish a assig me t method for distributi g tasks.If there were a ce tral admi istrator who k owswell about both the abilities of all IPUs a d the situatio s of all tasks, this problem would be fairly easy to solve.But i most cases, such a admi istrator does ot exist.Worse yet, IPUs ca ot k ow o e a others co ditio , the value of the best result, a d so forth.All what they k ow may o ly be:1, their ow curre t situatio s;a d 2, the feedback of the system after tasks are chose .Basi g o a simple hypothesis, that is, all IPUs have the same k owledge of the systems feedback, this paper proposes a task assig me t algorithm.Usi g some k owledge of probability theory, authors a alyze some performa ce aspects of this algorithm a d get their expressio s u der all situatio s.Some simplified forms of these expres- sio s are also give u der some situatio s usi g divide-a d-co quer a d extremum fi di g methods.I the ra ge of this paper, this i tegrated a alysis method is more simple a d practical tha Markov chai . Keywords　distributed systems;task assig me t;i depe de t processi g u it (IPU);self-lear i g;optima