Lecture 2

Lecture 2: Viewing a Complex System as an Open Network of Queues

\

For service systems, such as call centers and healthcare systems, just as for communication networks, computer systems and manufacturing systems, queueing models can be useful to understand, predict and improve system performance. In all these settings, it is useful to view these systems through the lens of queueing models.

  1. An Open Network of Queues

    • When considering the system broadly, the system as a whole can often be represented as an open network of queues. For this lecture, there is one main reading:

      • The Queueing Network Analyzer. by WW, Bell System Technical Journal, vol. 62, No. 9, November 1983, pp. 2779-2815.

          This paper is about a software tool to help do performance analysis of queueing systems. The main idea is that many systems can be fruitfully modeled as an open network of queues, as shown in Figure 1 on p. 2781. It can be very helpful when looking at systems and system data, to have this base model in mind. Key assumptions for the base model appear on p. 2781. For this perspective, we would also emphasize that we are assuming that the system can be considered stationary in time, so that the model is stationary. (The important issue of time-varying arrival rates will be addressed later.)

          In most complex systems of this kind, there are actually multiple classes of customers or jobs, each of which follows its own route or path through the network. In applications, we do not typically start with a model like Figure 1. Instead, we identify nodes or stations or queues and classes of customers or jobs that travel on routes or paths through the network. There is a process of aggregation to convert data by routes into the relatively simple network model in Figure 1. See Section 2.3; go trough Example 1 on p. 2791. After analyzing the simplified network in Figure 1, there also is a disaggregation step in obtaining performance descriptions for individual classes; e.g., to describe the expected time each class spends in the system before completing all required service. See Section 6.3.

  2. The Performance of a Single Queue

    • A key step in understanding the performance of a network of queues is to understand the performance of a single queue. Given the overall complexity, we should not require excessive mathematical precision. Our goal is to describe first-order performance. We will consider two main cases: (i) the single-server queue and (ii) the many-server queue. In both cases we will consider the FCFS service discipline, but it can be important to consider other service disciplines. For the single-server queue, we have a progression of models: M/M/1, M/GI/1, GI/GI/1, G/G/1. A similar progression exists for many-server queues. For service systems like call centers, it is also important to consider customer abandonment from queue.

      We provide the following brief overview of queueing theory as a guide to reading. References are given to basic introductions as well as additional advanced topics, which might be of interest to those familiar with the basic theory. We emphasize the importance of queueing models beyond the classical birth-and-death models, without the usual independence, Markov and exponential-distribution assumptions. (The need to go beyond stationary models will be discussed later.)

  3. Additional Topics and References

    • The QNA model features were subsequently expanded to address addtional complex realities of manufacturing systems, such as probabilistic transitions in the context of customer lass routes.

      • A Queueing Network Analyzer for Manufacturing. by Moshe Segal and WW, Teletraffic Science for New Cost-Effective Systems, Networks and Services, Proceedings of ITC 12 (ed. M. Bonatti), North-Holland, Amsterdam, 1989, pp. 1146-1152. (More complex algorithm to meet the needs of complex realities in manufacturing.)

    • Successful analysis can require properly accounting for dependence in point processes. This issue arises with superposition arrival processes.

    • Looking Ahead: Variability Impact in Many-Server Queues

      • Fluid Models for Multiserver Queues with Abandonments. by WW, Operations Research, vol. 54, No. 1, 2006, pp. 37-54 (For the steady-state performance of a stationary multi-server queues with abandonment, the service-time distribution beyond its mean does not matter much, but the patience-time distribution beyond its mean does matter; see Corollary 3.1 and Table 1.)

      • The G_t/GI/s_t+GI Many-Server Fluid Queue. by Yunan Liu (Columbia PhD, now at North Carolina State University) and WW, Submitted to Queueing Systems (For the transient behavior of a many-server queue, with or without abandonment, the service-time distribution beyond its mean does matter; see the example in Section 2, especially Figure 2.)

      • The Impact of Dependent Service Times on Large-Scale Service Systems. by Guodong Pang (Columbia PhD, now at Pennsylvania State University) and WW (with the aid of Andrew Li), To appear in Manufacturing and Service Operations Management. (It is possible to quantify the impact of dependence among the service times in a many-server queue, but is it important?)