Heavy traffic analysis of maximum pressure policies for stochastic processing networks with multiple bottlenecks (Q1025606)

The authors consider a stochastic processing network which takes inputs of various materials and uses different processing resources to produce output of various kinds. The control of such networks has found much research interest over the last years. Due to the complex structure it is rather convenient to investigate the networks' control under the assumption of heavy traffic where many details of the networks' microstructure disappear. This especially occurs in the situation of dominating bottlenecks as in the present paper. A common observation is then that a state space collapse occurs in the limit. Some details on the network's behavior: the connection between activities and resources, resp. activities and buffers is determined by incidence matrices. An activity may require simultaneous possession of several resources. The capacity of the resources can be split at any level, processing can be preempted. Within each buffer the head-of-the-line policy is applied. The stochastics of the network is determined by independent renewal sequences with finite (2+e) moment conditions and a routing matrix which controls the sequencing of jobs and activities. The overall control policy for the network is the maximum pressure policy which utilizes for decisions only the queue lengths at the serviceable buffer and the queue lengths at their immediately downstream buffers. From the literature it is known that under additional suitable conditions this policy is throughput optimal for general classes of processing networks. The authors contribute to this development by relaxing some of the side constraints. The results of the paper are centered around the machinery of proving heavy traffic limits for sequences of networks for diffusion scaled queue lengths and workload vectors to a semimartingale reflected Brownian motion living on a polyhedral cone.