ANOMALOUS RELATIONS AMONG VARIOUS PERFORMANCE OBJECTIVES IN DISTRIBUTED COMPUTER SYSTEMS 1

Hisao Kameda,Takayuki Kozawa,Jie Li

1997-01-01

Abstract: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.