Queueing theory

Queue networks are systems in which single queues are connected by a routing network. In this image, servers are represented by circles, queues by a series of rectangles and the routing network by arrows. In the study of queue networks one typically tries to obtain the equilibrium distribution of the network, although in many applications the study of the transient state is fundamental.

Queueing theory is the mathematical study of waiting lines, or queues. A queueing model is constructed so that queue lengths and waiting time can be predicted. Queueing theory is generally considered a branch of operations research because the results are often used when making business decisions about the resources needed to provide a service.

Queueing theory has its origins in research by Agner Krarup Erlang when he created models to describe the Copenhagen telephone exchange. The ideas have since seen applications including telecommunication, traffic engineering, computing and, particularly in industrial engineering, in the design of factories, shops, offices and hospitals, as well as in project management.

Spelling

The spelling "queueing" over "queuing" is typically encountered in the academic research field. In fact, one of the flagship journals of the profession is named Queueing Systems.

Single queueing nodes

Simple description and analogy

A queue, or "queueing node" can be thought of as nearly a black box. Jobs or "customers" arrive to the queue, possibly wait some time, take some time being processed, and then depart from the queue (see Fig. 1).

Fig. 1. A black box. Jobs arrive to, and depart from, the queue.

The queueing node is not quite a pure black box, however, since there is some information we need to specify about the inside of the queuing node. The queue has one or more "servers" which can each be paired with an arriving job until it departs, after which that server will be free to be paired with another arriving job (see Fig. 2).

Fig. 2. A queueing node with 3 servers. Server a is idle, and thus an arrival is given to it to process. Server b is currently busy and will take some time before it can complete service of its job. Server c has just completed service of a job and thus will be next to receive an arriving job.

An analogy often used is that of the cashier at a supermarket. There are other models, but this is one commonly encountered in the literature. Customers arrive, are processed by the cashier, and depart. Each cashier processes one customer at a time, and hence this is a queueing node with only one server. If a customer arrives to the cashier to find it busy, the customer will leave immediately. This is a queue with no buffer (or "waiting area", or any number of other terms). However, if up to 5 people may wait at the cashier while it is busy, this is a queue with a buffer of size 5.

Birth-death process

The behaviour / state of a single queue (also called a "queueing node") can be described by a Birth-death process, which describe the arrivals and departures from the queue, along with the number of jobs (also called "customers" or "requests", or any number of other things, depending on the field) currently in the system. An arrival increases the number of jobs by 1, and a departure (a job completing its service) decreases k by 1 (see Fig. 3).

Fig. 3. Birth / death process. The values in the circles represent the state of the birth-death process, k. The system transitions between values of k by "births" and "deaths" which occur at rates given by various values of λ_i and μ_i, respectively. For a queueing system, k is the number of jobs in the system (either being serviced or waiting if the queue has a buffer of waiting jobs). Further, for a queue, the arrival rates and departure rates are generally considered not to vary with the number of jobs in the queue so that we consider a single average rate of arrivals/departures per unit time to the queue. Hence, for a queue, this diagram has arrival rates of λ = λ₁, λ₂, …, λ_k and departure rates of μ = μ₁, μ₂, …, μ_k (see Fig. 4).

Fig. 4. A queue with 1 server, arrival rate λ and departure rate μ.

Kendall's notation

Single queueing nodes are usually described using Kendall's notation in the form A/S/c where A describes the distribution of durations between each arrival to the queue, S the distribution of service times for jobs and c the number of servers at the node. For an example of the notation, the M/M/1 queue is a simple model where a single server serves jobs that arrive according to a Poisson process (inter-arrival durations are exponentially distributed) and have exponentially distributed service times. In an M/G/1 queue, the G stands for general and indicates an arbitrary probability distribution for inter-arrival times.

Overview of the development of the theory

In 1909, Agner Krarup Erlang, a Danish engineer who worked for the Copenhagen Telephone Exchange, published the first paper on what would now be called queueing theory. He modeled the number of telephone calls arriving at an exchange by a Poisson process and solved the M/D/1 queue in 1917 and M/D/k queueing model in 1920. In Kendall's notation:

• M stands for Markov or memoryless and means arrivals occur according to a Poisson process;
• D stands for deterministic and means jobs arriving at the queue which require a fixed amount of service;
• k describes the number of servers at the queueing node (k = 1, 2, …).

If there are more jobs at the node than there are servers, then jobs will queue and wait for service

The M/G/1 queue was solved by Felix Pollaczek in 1930, a solution later recast in probabilistic terms by Aleksandr Khinchin and now known as the Pollaczek–Khinchine formula.

After the 1940s queueing theory became an area of research interest to mathematicians. In 1953 David George Kendall solved the GI/M/k queue and introduced the modern notation for queues, now known as Kendall's notation. In 1957 Pollaczek studied the GI/G/1 using an integral equation. John Kingman gave a formula for the mean waiting time in a G/G/1 queue: Kingman's formula.

Leonard Kleinrock worked on the application of queueing theory to message switching and packet switching. His initial contribution to this field was his doctoral thesis at the Massachusetts Institute of Technology in 1962, published in book form in 1964 in the field of digital message switching. His theoretical work after 1967 underpinned the use of packing switching in the ARPANET, a forerunner to the Internet.

The matrix geometric method and matrix analytic methods have allowed queues with phase-type distributed inter-arrival and service time distributions to be considered.

Problems such as performance metrics for the M/G/k queue remain an open problem.

Service disciplines

First in first out (FIFO) queue example.

Various scheduling policies can be used at queuing nodes:

First in first out

Also called first-come, first-served (FCFS), this principle states that customers are served one at a time and that the customer that has been waiting the longest is served first.

Last in first out

This principle also serves customers one at a time, but the customer with the shortest waiting time will be served first. Also known as a stack.

Processor sharing

Service capacity is shared equally between customers.

Priority

Customers with high priority are served first. Priority queues can be of two types, non-preemptive (where a job in service cannot be interrupted) and preemptive (where a job in service can be interrupted by a higher-priority job). No work is lost in either model.

Shortest job first

The next job to be served is the one with the smallest size

Preemptive shortest job first

The next job to be served is the one with the original smallest size

Shortest remaining processing time

The next job to serve is the one with the smallest remaining processing requirement.

Service facility

• Single server: customers line up and there is only one server
• Several parallel servers–Single queue: customers line up and there are several servers
• Several servers–Several queues: there are many counters and customers can decide going where to queue

Customer's behavior of waiting

• Balking: customers deciding not to join the queue if it is too long
• Jockeying: customers switch between queues if they think they will get served faster by doing so
• Reneging: customers leave the queue if they have waited too long for service

Arriving customers not served (either due to the queue having no buffer, or due to balking or reneging by the customer) are also known as dropouts and the average rate of dropouts is a significant parameter describing a queue.

Simple two-equation queue

A common basic queuing system is attributed to Erlang, and is a modification of Little's Law. Given an arrival rate λ, a dropout rate σ, and a departure rate μ, length of the queue L is defined as:

$L={\frac {\lambda -\sigma }{\mu }}.$

Assuming an exponential distribution for the rates, the waiting time W can be defined as the proportion of arrivals that are served. This is equal to the exponential survival rate of those who do not drop out over the waiting period, giving:

${\frac {\mu }{\lambda }}=e^{-W{\mu }}$

The second equation is commonly rewritten as:

$W={\frac {1}{\mu }}\mathrm {ln} {\frac {\lambda }{\mu }}$

The two-stage one-box model is common in epidemiology.

Queueing networks

Networks of queues are systems in which a number of queues are connected by what's known as customer routing. When a customer is serviced at one node it can join another node and queue for service, or leave the network.

For networks of m nodes, the state of the system can be described by an m–dimensional vector (x₁, x₂, …, x_m) where x_i represents the number of customers at each node.

The simplest non-trivial network of queues is called tandem queues. The first significant results in this area were Jackson networks, for which an efficient product-form stationary distribution exists and the mean value analysis which allows average metrics such as throughput and sojourn times to be computed. If the total number of customers in the network remains constant the network is called a closed network and has also been shown to have a product–form stationary distribution in the Gordon–Newell theorem. This result was extended to the BCMP network where a network with very general service time, regimes and customer routing is shown to also exhibit a product-form stationary distribution. The normalizing constant can be calculated with the Buzen's algorithm, proposed in 1973.

Networks of customers have also been investigated, Kelly networks where customers of different classes experience different priority levels at different service nodes. Another type of network are G-networks first proposed by Erol Gelenbe in 1993: these networks do not assume exponential time distributions like the classic Jackson Network.

Example analysis of an M/M/1 queue

Consider a queue with 1 server and the following variables:

• λ: the average arrival rate (customers arriving to the system per unit time, e.g. per 30 seconds);
• μ: the average departure rate (customers leaving the system—completing service—per the same unit time, e.g. per 30 seconds);
• n: the number of customers in the system at the given time;
• P_n: the probability of there being n customers in the system.

Further, let E_n represent the number of times the system enters state n, and L_n represent the number of times the system leaves state n. For all n, we have |E_n − L_n| ∈ {0, 1}, that is, the number of times the system leaves a state differs by at most 1 from the number of times it enters that state, since it will either return into that state at some time in the future (E_n = L_n) or not (|E_n − L_n| = 1).

When the system arrives at a steady state, the arrival rate should be equal to the departure rate.

Balance equation

situation 0: $\mu _{1}P_{1}=\lambda _{0}P_{0}$

situation 1: $\lambda _{0}P_{0}+\mu _{2}P_{2}=(\lambda _{1}+\mu _{1})P_{1}$

situation n: $\lambda _{n-1}P_{n-1}+\mu _{n+1}P_{n+1}=(\lambda _{n}+\mu _{n})P_{n}$

By balance equation, $P_{1}={\frac {\lambda _{0}}{\mu _{1}}}P_{0}\;\;\;P_{2}={\frac {\lambda _{1}}{\mu _{2}}}P_{1}+{\frac {1}{\mu _{2}}}(\mu _{1}P_{1}-\lambda _{0}P_{0})={\frac {\lambda _{1}}{\mu _{2}}}P_{1}={\frac {\lambda _{1}\lambda _{0}}{\mu _{2}\mu _{1}}}P_{0}$

By mathematical induction, $P_{n}={\frac {\lambda _{n-1}\lambda _{n-2}\cdots \lambda _{0}}{\mu _{n}\mu _{n-1}\cdots \mu _{1}}}P_{0}=P_{0}\prod _{i=0}^{n-1}{\frac {\lambda _{i}}{\mu _{i+1}}}$

Because $\sum _{n=0}^{\infty }P_{n}=P_{0}+P_{0}\sum _{n=1}^{\infty }\prod _{i=0}^{n-1}{\frac {\lambda _{i}}{\mu _{i+1}}}=1$

We get $P_{0}={\frac {1}{1+\sum _{n=1}^{\infty }\prod _{i=0}^{n-1}{\frac {\lambda _{i}}{\mu _{i+1}}}}}$

Routing algorithms

In discrete time networks where there is a constraint on which service nodes can be active at any time, the max-weight scheduling algorithm chooses a service policy to give optimal throughput in the case that each job visits only a single person service node. In the more general case where jobs can visit more than one node, backpressure routing gives optimal throughput. A network scheduler must choose a queuing algorithm, which affects the characteristics of the larger network. See also Stochastic scheduling for more about scheduling of queueing systems.

Mean field limits

Mean field models consider the limiting behaviour of the empirical measure (proportion of queues in different states) as the number of queues (m above) goes to infinity. The impact of other queues on any given queue in the network is approximated by a differential equation. The deterministic model converges to the same stationary distribution as the original model.

Fluid limits

Main article: fluid limit

Fluid models are continuous deterministic analogs of queueing networks obtained by taking the limit when the process is scaled in time and space, allowing heterogeneous objects. This scaled trajectory converges to a deterministic equation which allows the stability of the system to be proven. It is known that a queueing network can be stable, but have an unstable fluid limit.

Heavy traffic/diffusion approximations

Main article: heavy traffic approximation

In a system with high occupancy rates (utilisation near 1) a heavy traffic approximation can be used to approximate the queueing length process by a reflected Brownian motion, Ornstein–Uhlenbeck process or more general diffusion process. The number of dimensions of the RBM is equal to the number of queueing nodes and the diffusion is restricted to the non-negative orthant.

See also

• Ehrenfest model
• Erlang unit
• Industrial engineering
• Leonard Kleinrock
• Network simulation
• Poisson distribution
• Project Production Management
• Queue area
• Queueing delay
• Queue management system
• Queuing Rule of Thumb
• Queueing Systems – a journal of queueing theory
• Random early detection
• Renewal theory
• Throughput
• Scheduling (computing)
• Traffic jam
• Traffic generation model
• Flow network
• Stochastic scheduling

Further reading

• Gross, Donald; Carl M. Harris (1998). Fundamentals of Queueing Theory. Wiley. ISBN 978-0-471-32812-4. Online
• Deitel, Harvey M. (1984) [1982]. An introduction to operating systems (revisited first ed.). Addison-Wesley. p. 673. ISBN 978-0-201-14502-1. chap.15, pp. 380–412
• Kleinrock, Leonard (2 January 1975). Queueing Systems: Volume I – Theory. New York: Wiley Interscience. p. 417. ISBN 978-0471491101.
• Kleinrock, Leonard (22 April 1976). Queueing Systems: Volume II – Computer Applications. New York: Wiley Interscience. p. 576. ISBN 978-0471491118.
• Lazowska, Edward D.; John Zahorjan; G. Scott Graham; Kenneth C. Sevcik (1984). Quantitative System Performance: Computer System Analysis Using Queueing Network Models. Prentice-Hall, Inc. ISBN 978-0-13-746975-8.
• Zukerman, Moshe. Introduction to Queueing Theory and Stochastic Teletraffic Models (PDF).

External links

• Queueing theory calculator
• Teknomo's Queueing theory tutorial and calculators
• Office Fire Emergency Evacuation Simulation on YouTube
• Virtamo's Queueing Theory Course
• Myron Hlynka's Queueing Theory Page
• Queueing Theory Basics
• A free online tool to solve some classical queueing systems
• JMT: an open source graphical environment for queueing theory
• LINE: a general-purpose engine to solve queueing models
• What You Hate Most About Waiting in Line: (It’s not the length of the wait.), by Seth Stevenson, Slate, 2012 – popular introduction