Near-optimal load control of a message-passing multiprocessor system: A decentralized Markov decision approach

We address the problem of finding optimal load control policies for (call) control nodes in telecommunications networks. The node considered is a multiprocessing system with distributed memory in which the processing elements communicate through messages. The node can be modelled as an open queueing network with restricted class changes. Solving for a global state-dependent optimization policy would lead to intractable calculations. As an alternative approach, we show how the queueing network may be partitioned approximately into several isolated single processor queues, interconnected only through aggregates. Each single processor queue may in turn be formulated as a Markovian model, and by applying a Markov decision approach, the optimal regulator on each processor in isolation may be found. We propose a computational algorithm for solving this sub- optimal optimization problem and cross-check the results against a simpler steady-state flow model and simulations.