Imbedding gradient estimators in load balancing algorithms
Spiridon Pulidas, Don Towsley, John A. Stankovic · 2003
The problem of efficiently determining the optimum threshold parameter values for a decentralized load balancing algorithm is investigated. Simulation is used to study the behavior of a gradient-based decentralized optimization algorithm for obtaining good values. The algorithm computes the incremental job delay as a function of changes in both the local and remote job arrival rate. Estimators for these two quantities that are embedded in the optimization algorithm are described. Several experiments designed to evaluate the performance of the algorithm in a stationary environment and in an environment where there are changes in the workload are presented. The results indicate that the estimators are accurate, the algorithm chooses good thresholds, and the resultant response time of jobs is near optimal.>