Overall Optimal Load Balancing vs. Individually Optimal Load Balancing

Abstract

We can think of two load balancing policies each of which has a distinct performance objective. One is the overall optimal policy and the other is the individually optimal policy. The overall optimal load balancing policy is the policy whereby job scheduling is determined so as to minimize the system-wide mean job response time. We refer to the solution as the optimum. This policy has been studied by Tantawi and Towsley [TT85]. We also studied the policy in sections 1.1 and 1.2 this chapter. Of particular note, the conditions that the overall optimal solution satisfies and an efficient load balancing algorithm for multiple job classes are discussed. On the other hand, the individually optimal load balancing policy is the policy whereby each individual job is scheduled so as to minimize its own expected job response time, given the expected node and communication delays. We refer to the solution as the equilibrium. Wardrop considered this policy in traffic assignment problems [War52]. In the context of distributed computer systems, the policy has been studied by Kameda and Hazeyama [KH88].