Inferring queueing network models from high-precision location tracking data

Stochastic performance models are widely used to analyse the performance and reliability of systems that involve the flow and processing of customers. However, traditional methods of constructing a performance model are typically manual, time-consuming, intrusive and labourintensive. The limited amount and low quality of manually-collected data often lead to an inaccurate picture of customer flows and poor estimates of model parameters. Driven by advances in wireless sensor technologies, recent real-time location systems (RTLSs) enable the automatic, continuous and unintrusive collection of high-precision location tracking data, in both indoor and outdoor environment. This high-quality data provides an ideal basis for the construction of high-fidelity performance models. This thesis presents a four-stage data processing pipeline which takes as input high-precision location tracking data and automatically constructs a queueing network performance model approximating the underlying system. The first two stages transform raw location traces into high-level “event logs” recording when and for how long a customer entity requests service from a server entity. The third stage infers the customer flow structure and extracts samples of time delays involved in the system; including service time, customer interarrival time and customer travelling time. The fourth stage parameterises the service process and customer arrival process of the final output queueing network model. To collect large-enough location traces for the purpose of inference by conducting physical experiments is expensive, labour-intensive and time-consuming. We thus developed LocTrackJINQS, an open-source simulation library for constructing simulations with location awareness and generating synthetic location tracking data. Finally we examine the effectiveness of the data processing pipeline through four case studies based on both synthetic and real location tracking data. The results show that the methodology performs with moderate success in inferring multi-class queueing networks composed of singleserver queues with FIFO, LIFO and priority-based service disciplines; it is also capable of inferring different routing policies, including simple probabilistic routing, class-based routing and shortest-queue routing.