Predictive Load Balancing on Parallel Networks

Networks of workstations are commonly used alternatives to dedicated parallel machines, however one of their major drawbacks is the limitation on the communication bandwidth. In this paper, we use parallel network based techniques to optimize the available band-width. In real life, the data arrival and service rates are stochastic (dynamic) processes. Predictive filtering is a commonly used technique to estimate the future values of a stochastic process, based on its past values. We propose predictive filtering based load balancing algorithms for different stages of traffic handling. In the full version of this paper, based on the extensive simulation results, the proposed algorithms will be compared in terms of their performance, such as waiting time(average/maximum), interface utilization, and algorithm complexity.