Closed-loop load balancing: comparison of a discrete event simulation with experiments

Load balancing for parallel computations is modeled as a deterministic dynamic nonlinear time-delay system. This model accounts for the trade-off between using processor time/network bandwidth and the advantage of distributing the load evenly between the nodes to reduce overall processing time. A distributed closed-loop controller is presented to balance load dynamically at each node by using not only the local estimate of the queue size of other nodes, but also estimates of the number of tasks in transit. A discrete event simulation using OPNET Modeler is presented and compared with experimental data, and results indicate good agreement between the nonlinear time-delay model and the behaviors observed on a parallel computer network. Moreover, both simulations and experiments show a dramatic increase in performance obtained using the proposed closed-loop controller.