| Stochastic multiclass queueing networks (MQNETs) can be used to model a plethora of practical systems, including manufacturing systems, communication networks, multiprocessor computer systems and networks of computing servers. Performance in these systems is drastically affected by the sequence in which jobs are processed at various nodes (scheduling) and the way jobs are routed through the network (routing). Performance analysis in MQNETs is notoriously hard. Naturally, control, in the form of scheduling or routing, is even harder.; We propose a new parametric class of scheduling and routing policies for open and closed MQNETs, which we call target-pursuing policies. In open networks, they steer the state of the system towards a pre-determined and fixed target, while, in closed networks they steer instantaneous throughputs towards a fixed target. In both cases, the proposed policies measure distance from the target using a weighted norm. In both open and closed networks, the proposed policies are amenable to distributed implementation using local state information.; In open networks for certain choices of the norm we show that target-pursuing policies are stable, i.e., the number of jobs in the MQNET remains finite at all times.; To that end, we work with an appropriate fluid model of the system. In closed networks, we prove that with proper target selection the corresponding policy is efficient, that is, attains bottleneck throughput in the infinite population limit.; The target-pursuing policies we propose depend on a number of parameters. We outline how to select appropriate policy parameter values to optimize performance. A rich set of numerical results indicates that our approach leads to near-optimal policies (when the optimal can be computed) or policies that significantly outperform heuristic alternatives. |
