Queueing and Service-System Analytics

Introduction

Queueing systems arise whenever demand for a service is not perfectly synchronized with service capacity. Customers may be patients, telephone calls, jobs in a cloud platform, trucks at a loading dock, or claims awaiting review. Congestion is not simply an inconvenience: it affects revenue, customer experience, safety, employee workload, inventory, and the cost of capacity.

Queueing theory provides a disciplined way to analyze these systems under uncertainty. A useful model links three elements: the stochastic arrival process, the service mechanism, and the rule governing service order. The analyst then estimates measures such as utilization, average queue length, waiting time, probability of delay, and the capacity required to achieve a service-level target.

The models in this chapter are deliberately idealized. Their value lies not in predicting every individual arrival, but in clarifying the structural relationship between variability, utilization, capacity, and delay. Graduate-level analysis requires both correct calculation and critical evaluation of whether model assumptions represent the operating environment.

Learning Objectives

By the end of this chapter, you should be able to:

1. Describe a service system using Kendall notation.
2. Formulate and interpret birth–death queueing processes.
3. Apply Little’s Law and distinguish time-average from customer-average measures.
4. Calculate performance measures for M/M/1 and M/M/c systems.
5. Analyze finite-capacity and priority queues.
6. Explain queueing networks and the role of service-level measures.
7. Evaluate staffing and capacity trade-offs under uncertainty.
8. Identify limitations, governance concerns, and appropriate uses of simulation and data-driven methods.

1. Anatomy of a Queueing System

A queueing system consists of a population of potential customers, an arrival process, one or more queues, servers, and a service discipline. The queue may be physical or virtual. A call-center queue, for example, may hold calls in software while agents are the servers.

Important terminology includes:

• Arrival rate (λ): expected arrivals per unit of time.
• Service rate (μ): expected completions per server per unit of time.
• Utilization (ρ): the fraction of capacity demanded, subject to the model definition.
• Waiting time: time spent before service begins.
• Sojourn or system time: waiting time plus service time.
• Queue length: number waiting, excluding customers in service.
• System size: number waiting plus number in service.
• Service discipline: rule determining which customer is served next.

Common disciplines include first-come, first-served (FCFS), last-come, first-served, random selection, shortest processing time, and priority classes. FCFS is often viewed as fair, but it is not necessarily optimal for minimizing average waiting time. Shortest-processing-time rules can reduce mean delay when processing times are known, while priority rules may protect urgent work at the expense of lower-priority customers.

Kendall Notation

A queue is commonly summarized as A/S/c/K/N/D, where:

• A describes the arrival-time distribution;
• S describes the service-time distribution;
• c is the number of parallel servers;
• K is system capacity, including customers in service;
• N is the calling-population size; and
• D is the service discipline.

The symbols M, D, and G mean Markovian (exponential), deterministic, and general distributions. If later components are omitted, conventional defaults are usually infinite capacity, an infinite calling population, and FCFS. Thus M/M/1 generally means Poisson arrivals, exponential service, one server, unlimited capacity, and FCFS.

Kendall notation is a modeling language, not a complete description. Two M/M/1 systems can have different customer classes, abandonment rules, or time-varying rates and therefore different practical behavior.

2. Birth–Death Processes and Stability

A birth–death process represents a system whose state changes by one customer at a time. State n is the number in the system. A birth moves from n to n+1 and an associated death moves from n to n−1. For an M/M/1 queue, births occur at rate λ and deaths occur at rate μ whenever the system is nonempty.

For a general birth–death process with state-dependent birth rates λ_n and death rates μ_n, the stationary probabilities satisfy the local balance relationship

where π_n is the long-run probability of being in state n. Consequently,

with π₀ selected so that ∑_n=0^∞ π_n = 1. The formula applies when a stationary distribution exists and rates are properly defined.

For M/M/1, define utilization as

Here both λ and μ have units of customers per unit time, so ρ is dimensionless. The unlimited-capacity system is stable only when ρ < 1. If ρ ≥ 1, the long-run average number in the system is not finite under the basic model. Temporary operation may still be possible, but congestion tends to grow rather than return to a stable distribution.

The stationary probability of n customers is

M/M/1 Performance Measures

Under Poisson arrivals, exponential service, one server, infinite capacity, stability, and FCFS:

Here L is the mean number in the system, L_q is the mean number waiting, W is mean system time, and W_q is mean waiting time. The time units of W and W_q are the reciprocal of the rate units.

The identity W = W_q + 1/μ follows because exponential service has mean duration 1/μ. As ρ approaches one, waiting increases sharply. This nonlinear effect explains why operating a service system at nominal full capacity can produce unacceptable delays.

Worked Example: Single Analyst

An analytics help desk receives λ = 12 requests per hour. One analyst completes requests at μ = 15 per hour. Assuming M/M/1 conditions,

The analyst is busy 80% of the time, but an arriving request experiences an average total time of 20 minutes. The result is an average, not a guarantee: individual waits vary substantially.

3. Little’s Law

Little’s Law is one of the most robust relationships in operations analytics:

For a queue alone,

L is the time-average number in the relevant system, λ_eff is the effective throughput rate, and W is the average time spent by customers in that system. Units are consistent: customers = customers/time × time.

The law does not require Poisson arrivals, exponential service, or a particular queue discipline. It requires a stable system and consistent definitions over a sufficiently long observation period. In a finite-capacity queue, λ_eff is the rate admitted into service, not necessarily the external attempted-arrival rate. In systems with abandonment, it must be clear whether abandoned customers are included.

Little’s Law is particularly useful when direct measurement of one quantity is difficult. If a claims operation contains an average of 240 open claims and processes 30 claims per day, the implied average time in the process is W = 240/30 = 8 days, provided the population and throughput boundaries match.

4. Multi-Server Queues: M/M/c

An M/M/c queue has Poisson arrivals, exponential service times, and c identical parallel servers. The total service-completion rate in state n is

The stability condition for unlimited capacity is

It is also useful to define the offered load a = λ/μ, measured in server-equivalents. The probability that the system is empty is

The probability that an arriving customer must wait, known as the Erlang C delay probability, is

The expected queue length and waiting time are

These expressions assume unlimited capacity, identical independent servers, FCFS, stationary rates, and no abandonment.

Worked Example: Staffing a Support Desk

A support desk receives 18 cases per hour. Each agent handles 10 cases per hour on average. With two agents, ρ = 18/(2×10) = 0.9, a high utilization level. With three agents, ρ = 18/(3×10) = 0.6. For c=3, a=18/10=1.8.

Thus approximately 38.4% of arrivals wait. Further,

and

The third agent substantially reduces delay, but its economic value depends on labor cost, the value of timely service, and the target service level.

5. Finite-Capacity Queues and Loss Systems

In an M/M/1/K queue, at most K customers can be in the system. An arrival finding K customers is blocked, lost, or redirected. For ρ ≠ 1,

For ρ=1, all states have probability 1/(K+1). The blocking probability is P_block=π_K, and the effective arrival rate is

The relation between effective throughput and system time is therefore L=λ_eff W, not necessarily L=λW.

A loss system with c servers and no waiting room is often modeled as M/M/c/c. Its blocking probability is given by the Erlang B formula:

where a=λ/μ. This model is relevant to telephone trunks, finite parking spaces, and some computing resources. Blocking is not automatically bad: in a deliberately engineered loss system, rejecting excess demand may be preferable to creating long waits.

6. Priority Queues

Priority systems classify customers and select the next customer according to class. Priorities may be strict or non-strict, preemptive or non-preemptive, and static or dynamic. In a preemptive-resume system, a high-priority arrival can interrupt a lower-priority job, which later resumes. In a non-preemptive system, service already in progress continues.

Priority improves response for favored classes but creates externalities. Lower-priority work may experience very long waits or starvation. A fair design therefore needs explicit policies, monitoring, and escalation rules.

For an M/G/1 non-preemptive priority queue with classes ordered from highest to lowest, let λ_i be class i arrival rate, E[S_i] its mean service time, and ρ_i=λ_iE[S_i]. Define ρ_<i=∑_j=1ⁱ⁻¹ρ_j. Under standard independence and stability assumptions, the mean waiting time for class i is

where

The second moment E[S_j²] captures service-time variability. This formula illustrates two important facts: variability creates delay, and priority redistributes delay across classes. It should not be applied without checking the precise priority discipline and assumptions.

7. Queueing Networks

A queueing network contains multiple service stations connected by routing. In an open network, customers can enter from outside and eventually leave. In a closed network, a fixed population circulates among stations, as in a computer system with a fixed number of users.

For an open network, the traffic equations are

where γ_j is the external arrival rate to node j, p_ij is the probability of routing from node i to node j, and λ_j is the total arrival rate to node j. Solving these simultaneous equations identifies bottleneck stations.

For example, if a workflow receives 100 jobs per hour, 70% visit validation, and every validated job enters fulfillment, then fulfillment’s rate is 70 jobs per hour if 70% is the relevant routing probability. A feedback loop can make internal rates substantially greater than external demand.

Jackson network results provide product-form stationary distributions under restrictive conditions, including Poisson external arrivals, exponential service, suitable independent routing, and stable nodes. Real networks often violate these assumptions through correlated arrivals, batching, blocking, finite buffers, and rework. Simulation or more specialized analysis may then be necessary.

8. Service Levels, Staffing, and Capacity Trade-Offs

Mean waiting time is not sufficient for managerial decisions. Customers often care about a percentile or a threshold probability. Common measures include:

• probability of waiting: P(W_q>0);
• probability of answering within target t: P(W_q≤t);
• probability of abandonment;
• fraction completed within a deadline;
• fill rate or on-time completion rate; and
• maximum or percentile queue length.

For an M/M/1 queue, the waiting-time distribution has a point mass at zero and, conditional on waiting, an exponential tail. Under the standard model,

Thus a target P(W_q≤t)≥α requires

This equation is model-specific and should not be confused with a general service-level law.

Capacity decisions balance labor or infrastructure cost against delay, abandonment, lost sales, quality risk, and reputational damage. Because congestion rises nonlinearly with utilization, a small amount of spare capacity may have high option value during demand surges. However, permanent overstaffing can be wasteful if demand is predictable and flexible staffing is available.

A practical staffing process is:

1. Forecast demand by interval rather than relying only on daily averages.
2. Estimate service-time distributions and shrinkage, including breaks, training, and absence.
3. Select a queueing approximation or simulation model.
4. Test staffing alternatives against service-level and cost targets.
5. Stress-test demand surges, absenteeism, and parameter uncertainty.
6. Monitor actual performance and recalibrate.

Pooling often improves efficiency because variability is shared across servers. Yet pooling can increase walking distance, specialization loss, or coordination cost. Capacity is therefore an operational design choice, not merely a numerical ratio.

9. Illustrative Case: A Digital Claims Center

Background

A fictional insurer, Northbridge Mutual, receives digital claims throughout the day. Claims are screened by a general intake team and then routed to specialist units. Customers expect 90% of simple claims to receive an initial decision within 30 minutes.

Problem

The center staffs to the average daily arrival rate. During lunch and weather-related surges, the queue grows rapidly. Management proposes either adding permanent staff or introducing an urgent-claim priority class.

Analysis

The analyst first segments arrivals by 15-minute interval and estimates service-time distributions. The average obscures a peak arrival rate of 48 claims per hour, while each specialist processes 12 claims per hour. Four specialists would therefore have nominal utilization ρ=48/(4×12)=1, which is unstable in an unlimited-capacity model. Five specialists yield ρ=0.8, but the M/M/5 calculation still predicts a nonzero probability of delay. Historical data also show that service times are more variable than exponential and that 6% of customers abandon after extended waits.

A simulation is built with time-varying arrivals, empirical service times, abandonment, breaks, and routing. Scenarios compare four, five, and six specialists; cross-training is included as a staffing alternative. A priority rule for urgent claims is tested with a cap limiting urgent work to 25% of capacity.

Approach and Outcome

The simulation indicates that five specialists meet the average target but fail during two-hour surges. Six specialists meet the 90% target more consistently, while cross-training one intake employee during peaks achieves similar service performance at lower expected cost. Unrestricted urgent priority improves urgent response but causes unacceptable delays for ordinary claims. The capped priority rule provides improvement without severe starvation.

Learning Points

The case illustrates why average utilization is not a service-level guarantee, why time-varying demand matters, and why priority rules should be evaluated by their effect on every class. It also shows that simulation is valuable when exponential and stationary assumptions are implausible, but the model must be validated against observed arrivals, service times, routing, and abandonment.

10. Limitations, Governance, and Contemporary Analytics

Queueing models can mislead when arrival rates vary rapidly, service times are heavy-tailed, customers abandon, servers are heterogeneous, or work is batched. Dependence between arrivals and service times can invalidate simple formulas. Estimates should therefore include confidence intervals or sensitivity ranges where possible.

AI and machine learning can forecast arrivals, classify work, estimate service times, detect anomalies, and recommend staffing. Forecast accuracy does not remove queueing risk: an unbiased forecast can still generate poor service if variability, forecast error, or operational constraints are ignored. Prescriptive systems should be monitored for drift and tested under stress.

Governance is especially important when queues allocate access to healthcare, credit review, public services, or employment processes. Priority rules may reproduce historical inequities if urgent status is correlated with protected characteristics or if data reflect unequal access. Decision rules should be documented, auditable, explainable to affected users, and reviewed for disparate impact. Privacy controls are required when individual-level arrival, location, health, or behavioral data are used.

Best practice is to combine analytical formulas, operational data, simulation, and managerial judgment. A model should state its boundary, assumptions, parameter sources, validation method, and decision consequences. The objective is not to maximize utilization; it is to deliver reliable service at an economically and ethically defensible level.

Chapter Summary

Queueing and service-system analytics connect uncertain demand with capacity and delay. Kendall notation identifies arrival, service, server, capacity, population, and discipline assumptions. Birth–death processes provide a foundation for stationary queue probabilities, while M/M/1 formulas show the nonlinear growth of congestion as utilization approaches one. Little’s Law links average inventory, throughput, and time across a wide range of systems.

M/M/c models support multi-server staffing decisions; finite-capacity models account for blocking and effective throughput; priority models reveal the trade-off between urgent response and fairness; and queueing networks expose routing and bottlenecks. Service levels require more than mean delay, often emphasizing percentiles or deadline compliance. Real decisions should incorporate time-varying demand, variability, abandonment, cost, ethics, and model uncertainty. Analytical formulas are powerful benchmarks, while simulation and empirical validation are essential when assumptions are unrealistic.

Key Terms

Kendall notation: Standard notation describing arrival process, service process, servers, capacity, population, and discipline.

Arrival rate: Expected number of attempted arrivals per unit time.

Service rate: Expected completion rate of one server.

Utilization: Fraction of available service capacity demanded.

Birth–death process: Stochastic process whose state changes by one upward or downward transition.

Stationary distribution: Long-run probabilities of occupying each system state.

Little’s Law: Relationship L=λ_effW between average number, throughput, and time.

M/M/1: Poisson arrivals, exponential service, and one server.

M/M/c: Poisson arrivals, exponential service, and c parallel servers.

Erlang C: Formula for delay probability in an M/M/c queue with waiting room.

Erlang B: Formula for blocking probability in a loss system.

Finite capacity: Queueing system with a maximum number of customers present.

Priority discipline: Rule that ranks classes for service selection.

Preemption: Interruption of one service activity to serve another.

Queueing network: Connected collection of queues with routing between stations.

Bottleneck: Station whose capacity most constrains network throughput.

Service level: Probability or proportion meeting a specified performance target.

Abandonment: Customer departure before service begins or completion.

Offered load: Arrival rate divided by mean service rate, measured in server-equivalents.

Review Questions

Short Answer Questions

1. What information is conveyed by Kendall notation M/M/1/K?
2. Why must utilization be less than one for a stable unlimited-capacity M/M/1 queue?
3. State Little’s Law and explain each term.
4. Distinguish Erlang B from Erlang C.
5. Why can a priority rule improve one class’s waiting time while worsening another’s?

Long Answer Questions

1. Derive the stationary probabilities for an M/M/1 queue using local balance and normalization.
2. Explain the assumptions behind M/M/c performance formulas and discuss situations in which they may fail.
3. Compare finite-capacity, infinite-capacity, and loss-system designs from an operational perspective.
4. Explain how routing equations identify bottlenecks in an open queueing network.
5. Develop a staffing framework that combines cost, utilization, service level, abandonment, and demand uncertainty.

Analytical / Application Questions

1. A clinic receives 10 patients per hour and one physician serves 12 per hour. Using an M/M/1 model, calculate utilization, mean queue length, and mean waiting time. Interpret the operational risk.
2. A service center receives 24 jobs per hour, and each server processes 10 jobs per hour. Compare the utilization and likely delay implications of two versus three servers. What additional data would you require before recommending staffing?
3. Design an evaluation plan for a priority queue in a public-service organization. Specify performance measures for high- and low-priority customers, fairness checks, data requirements, and stress tests.

Multiple Choice Questions with Answers

Question 1

For an M/M/1 queue with arrival rate λ=8 per hour and service rate μ=10 per hour, what is utilization?

A. 0.20

B. 0.80

C. 1.25

D. 2.00

Correct Answer: B. 0.80

Question 2

Which condition is required for a stationary unlimited-capacity M/M/1 queue?

A. λ>μ

B. λ=μ

C. λ<μ

D. λ=0

Correct Answer: C. λ<μ

Question 3

A process has an average of 150 cases in progress and throughput of 25 cases per day. By Little’s Law, what is the average time in the process?

A. 4 days

B. 6 days

C. 25 days

D. 175 days

Correct Answer: B. 6 days

Question 4

In an M/M/c queue, what does Erlang C estimate?

A. The probability that an arrival must wait

B. The probability that all arrivals are rejected

C. The average service time only

D. The number of routing destinations

Correct Answer: A. The probability that an arrival must wait

Question 5

Which model is most appropriate when a customer finding a full system is blocked and there is no waiting room?

A. M/M/1 with infinite capacity

B. M/M/c with infinite waiting room

C. M/M/c/c loss system

D. Deterministic inventory model

Correct Answer: C. M/M/c/c loss system

Question 6

In a finite-capacity queue, effective arrival rate is generally

A. the attempted arrival rate multiplied by the admission probability

B. always equal to the service rate

C. always greater than the attempted arrival rate

D. independent of blocking

Correct Answer: A. the attempted arrival rate multiplied by the admission probability

Question 7

What is the principal operational danger of operating an unlimited-capacity queue near utilization one?

A. Service times become deterministic

B. Mean delay can increase sharply with small increases in demand

C. All customers are served immediately

D. The queue becomes independent of arrivals

Correct Answer: B. Mean delay can increase sharply with small increases in demand

Question 8

In an open queueing network, traffic equations are primarily used to determine

A. customer identities

B. total arrival rates at network nodes

C. employee salaries

D. the exact completion time of every job

Correct Answer: B. total arrival rates at network nodes

Question 9

Which statement about Little’s Law is correct?

A. It requires exponential service times in every application

B. It applies only to one-server queues

C. It links average number, effective throughput, and average time under consistent system boundaries

D. It assumes zero variability

Correct Answer: C. It links average number, effective throughput, and average time under consistent system boundaries

Question 10

Why should priority queues be evaluated using measures for all customer classes?

A. Priority never changes waiting times

B. Improving one class can impose substantial delay or starvation on another

C. Only low-priority customers generate arrivals

D. Priority makes capacity unlimited

Correct Answer: B. Improving one class can impose substantial delay or starvation on another