ANOMALOUS RELATIONS AMONG VARIOUS PERFORMANCE OBJECTIVES IN DISTRIBUTED COMPUTER SYSTEMS 1
ANOMALOUS RELATIONS AMONG VARIOUS PERFORMANCE OBJECTIVES IN DISTRIBUTED COMPUTER SYSTEMS 1
复制标题
分布式计算机系统中各种性能目标之间的异常关系1
DOI：
--
复制
发表时间：
1997
期刊：
影响因子：
--
通讯作者：
Jie Li
中科院分区：
文献类型：
--
作者：
H. Kameda;T. Kozawa;Jie Li
分布式计算机系统由一组由通信网络连接的异构主机（即节点）组成。到达一个节点的作业可以在本地处理，也可以转移到另一个节点进行远程处理，我们称之为负载平衡。分布式计算机系统中负载平衡的一个可能的性能目标是最小化总体平均响应时间。我们可以解析地描述静态负载平衡策略，其中平均总体响应时间最小，我们称之为总体最优策略。然而，这个策略在某种意义上缺乏公平性，例如，到达同一节点但被转发到不同节点的两个作业可能没有相同的预期响应时间。为了满足作业之间的公平性，我们可以考虑一个单独的最优负载平衡策略，即到达同一节点的作业具有相同的（最小）预期响应时间，而不管处理它们的节点是什么。此外，我们可以考虑一个节点最优负载平衡策略，即到达每个节点的作业的平均响应时间最小，因为到达这些节点的作业被转发的决策是由其他节点决定的。我们报告了上述政策之间的相互关系中存在一些看似反常的现象。
Distributed computer systems consist of a set of heterogeneous host computers (i.e., nodes) connected by a communication network. A job that arrives at a node may either be processed locally or transferred to another node for remote processing, which we call load balancing. One possible performance objective of load balancing in distributed computer systems is to minimize the overall mean response time. We can characterize analytically the static load balancing policy whereby the mean overall response time is minimized, which we call the overall optimal policy. This policy, however, lacks fairness in the sense that, for example, two jobs arriving at the same node but being forwarded to different nodes may not have the same expected response time. To satisfy fairness among jobs we can consider an individually optimal load balancing policy whereby jobs arriving at the same node have the same (minimum) expected response time regardless of the nodes which process them. Furthermore, we can think of a node optimal load balancing policy whereby the mean response time of jobs arriving at each node is minimum given the decision by the other nodes of which jobs arriving at those nodes are forwarded. We report the existence of some seemingly anomalous phenomena in the mutual relation among the above policies.