A processing network is a system that takes materials of various kinds as inputs, and uses processing resources to produce other materials as outputs. A wide range of real world, complex systems can be represented within this conceptual framework, including semiconductor wafer fabrication facilities, networks of data switches, and large-scale call centers. In a manufacturing context, the processing network paradigm is flexible enough to accommodate machine-operator interactions, material handling equipment, machine breakdown, and assembly-disassembly operations. It can be used to model call centers with cross-trained operators, to model input-queued data switches, and to model congestion-based routing of road traffic. Key performance measures of a stochastic processing network include throughput (the rate at which jobs or materials flow through the system) and cycle time (the total amount of time that inputs spend in the network). Elements of an operational policy may include input control, sequencing, routing, and decisions related to batching and setups; the choice of such a policy can dramatically affect network performance.

In this talk, it will first be shown that even in simple networks, commonly used operational policies like first-in-first-out sequencing may perform badly, failing to achieve even "throughput optimality." We then present two families policies, called discrete-proportional-processor-sharing and maximum pressure policies, that are always throughput optimal, regardless of the processing network's topology or parameter values. These policies have other attractive features as well, including distributed implementation that uses only local or semi-local congestion information. A simulation study has been undertaken that evaluates these policies in a wafer fabrication setting, using SEMATECH data sets. The results of that study will be discussed, along with other attractive theoretical properties of the two policy families.