Adaptive load balancing in a distributed environment

We propose an architecture for an embedded adaptive scheduler in a heterogeneous workstation network. The generic architecture is applicable to various balancing problems arising in a distributed environment. As an example we introduce an adaptive job scheduler. The scheduler gives recommendations for a non-preemptive job transfer between the participating workstations. A neural network algorithm is used to improve the knowledge of the scheduler by learning from the previous behaviour of the job. The scheduler adapts very quickly to various jobs as well as to the changing environment, whereby the calculation overhead is negligible. Results from a prototype implementation demonstrate the behaviour of the scheduler and the performance benefit for the system