Abstract
Clusters of workstations is a commonly used alternative to dedicated parallel machines, however one of their major drawbacks is the communication bandwidth. One way to overcome this difficulty is to use parallel networks [1, 2, 3, 4, 5, 6, 7, 8, 9, 16], where each workstation has possibly more than one network interface unit and connected to each by more than one network. In this paper, we consider the traffic balancing problem on parallel networks, namely how to utilize the available network resources to reduce the communication delay in the best possible way. One of the major challanges in this area is that the network traffic is stochastic and is not easy to predict accurately. In the case of perfect prediction, one can find the optimum utilization policy by solving a simple linear programming problem. If perfect predictions are not available or possible, one may first try to randomly distribute the load to the existing parallel networks, however by using stochastic estimation techniques one can achieve much better utilization of the parallel network compared to the random assignment approach. To achieve this, we propose an autoregressive model based prediction method and use it together with some linear programming techniques to find the optimum traffic balancing policy for the parallel network.