Variability, Queuing, Bottlenecks, and Flow Control

Introduction and Context

Capacity planning is often presented as a comparison between available capacity and expected demand. In practice, however, flow systems rarely operate under stable, perfectly predictable conditions. Arrivals vary by time of day and day of week; processing times differ across jobs, employees, and machines; equipment fails; materials arrive late; and priorities change. These sources of variability mean that a system with nominal capacity equal to demand may perform poorly in practice.

Queuing theory explains how variability creates waiting and congestion. Bottleneck analysis identifies the resource that most restricts throughput. Little’s Law connects work-in-process, throughput, and flow time. Finite-buffer analysis explains why blocking and starving occur in production lines. The Theory of Constraints (TOC) provides a managerial method for exploiting and subordinating the constraint before expanding it. Together, these concepts support decisions about staffing, batch sizes, buffers, schedules, maintenance, line balancing, and customer promises.

The central managerial insight is that utilization, variability, and waiting are interdependent. High utilization is attractive because idle capacity appears wasteful, but as utilization approaches 100 percent, even modest variability can cause disproportionately large queues and delays. Effective flow control therefore seeks not merely to maximize local utilization but to maximize reliable system throughput.

Learning Objectives

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

1. Explain how arrival and processing variability affect queues and flow time.
2. Apply basic queuing models and interpret their assumptions and limitations.
3. Use Little’s Law to diagnose work-in-process and flow-time problems.
4. Identify bottlenecks and distinguish bottleneck utilization from nonbottleneck utilization.
5. Apply the five focusing steps of the Theory of Constraints.
6. Analyze blocking, starving, and finite-buffer effects in production lines.
7. Recommend operational interventions that protect bottleneck capacity and improve flow.
8. Evaluate the limitations, governance implications, and technology requirements of flow-control systems.

1. Variability as a Capacity Problem

Variability may be classified according to its source and its managerial controllability.

A useful distinction is between natural variability, which is inherent to the process or demand, and induced variability, which is created by management policies such as large batches, unstable priorities, excessive expediting, or frequent schedule changes.

Variability affects capacity in two ways. First, it creates periods in which demand exceeds instantaneous service capacity, producing queues. Second, it creates idle periods when a resource lacks work. These effects do not cancel each other from a customer’s perspective: idle time cannot usually be recovered to eliminate the waiting caused during a peak period.

For a process with mean processing time E[S] and service rate μ, the nominal capacity of one server is:

where μ is measured in jobs per unit of time and E[S] is the mean service time in the same time unit. This expression assumes a continuously available server and excludes downtime, setups, quality losses, and other capacity reductions. Practical capacity is therefore lower than nominal capacity.

2. Queuing Fundamentals

A queue consists of entities waiting for service plus the service facility itself. In operations management, entities may be customers, orders, patients, materials, or jobs. A queuing system is commonly described by:

• the arrival process;
• the number of servers and their configuration;
• the service-time distribution;
• queue discipline, such as first-come-first-served or priority service;
• system capacity, including any finite waiting space; and
• the population from which arrivals originate.

The notation A/S/c/K often summarizes a queue, where A describes the arrival distribution, S the service-time distribution, c the number of parallel servers, and K the maximum number of entities in the system. For example, M/M/1 denotes Poisson arrivals, exponentially distributed service times, and one server. The letter M indicates a memoryless distribution.

Utilization and Stability

For a single-server system, utilization is:

where λ is the mean arrival rate, μ is the service rate, and ρ is dimensionless. For a stable infinite-capacity queue, the long-run arrival rate must be less than the service rate, so ρ < 1. For c identical parallel servers, utilization is commonly expressed as:

This is a necessary stability condition, not a guarantee of short waits. Variability and the chosen service discipline still determine performance.

The M/M/1 Model

For an M/M/1 queue with Poisson arrivals, exponential service times, one server, unlimited waiting space, independent arrivals and services, and first-come-first-served operation, the mean number in the system is:

The mean time in the system is:

The mean number waiting and mean waiting time before service are:

Here L and Lq are entities, while W and Wq are units of time. These relationships are special results for the M/M/1 model and should not be applied automatically to general service systems.

Worked Example 1: Utilization and Waiting

A repair technician receives an average of 6 jobs per hour. Mean repair time is 8 minutes, or 0.1333 hours. The service rate is therefore:

μ = 1/0.1333 ≈ 7.5 jobs/hour

Utilization is:

ρ = 6/7.5 = 0.8

Using the M/M/1 formulas:

L = 0.8/(1−0.8) = 4 jobs

W = 1/(7.5−6) = 0.6667 hours = 40 minutes

L_q = 0.8²/(1−0.8) = 3.2 jobs

W_q = 0.8/(7.5−6) = 0.5333 hours = 32 minutes

Thus, although the average repair itself takes only 8 minutes, the expected total time in the system is 40 minutes. The example illustrates the nonlinear effect of utilization: increasing utilization from 80 percent toward 100 percent causes waiting to rise sharply.

3. General Variability and the VUT Relationship

Many operational processes are not M/M/1 systems. Service times may be deterministic, highly variable, or affected by setups and rework. A useful approximation for a single-server queue is the VUT relationship:

where:

• Wq is approximate mean waiting time in queue;
• u is utilization;
• Ca is the coefficient of variation of interarrival times;
• Cs is the coefficient of variation of service times; and
• te is mean effective processing time.

The coefficient of variation is:

where σ is the standard deviation and μx is the mean of the measured variable. C is dimensionless. The VUT relationship is an approximation, especially useful for diagnosing how utilization and variability jointly influence waiting. It assumes a reasonably stable single-server setting and should not be treated as an exact result for finite buffers, complex networks, priorities, or strongly dependent arrivals.

The expression highlights three managerial levers:

1. Variability reduction: reduce Ca through appointment systems, workload leveling, and controlled release; reduce Cs through standard work, preventive maintenance, and quality improvement.
2. Utilization management: protect practical capacity by avoiding sustained operation near 100 percent.
3. Time reduction: reduce mean effective processing time through better methods, automation, layout, and setup reduction.

A key implication is that reducing variability can increase effective capacity even when no machine or employee is added.

4. Little’s Law and Flow Diagnosis

Little’s Law is a general steady-state relationship:

where L is the average number of entities in the system, λ is the long-run average throughput rate, and W is the average time an entity spends in the system. The corresponding queue-only relationship is:

Little’s Law does not require Poisson arrivals, exponential service, or a particular queue discipline. It requires a stable system observed over a sufficiently long period, with consistent definitions of the system boundary and the measured averages. Units must be consistent: jobs = jobs/hour × hours.

Worked Example 2: Diagnosing Work-in-Process

A claims department completes 120 claims per day and has an average of 600 claims in process. Little’s Law gives:

W = L/λ = 600/120 = 5 days

If management wants average flow time of 3 days at the same throughput, target work-in-process is:

L = λW = 120×3 = 360 claims

Therefore, 240 claims must be removed from the system on average, either by reducing input, increasing effective throughput, eliminating rework, or redesigning the process. Simply instructing employees to “work faster” without addressing releases and bottlenecks is unlikely to achieve the target.

Little’s Law is also useful for identifying inconsistent performance reports. For example, if a manager reports 1,000 jobs in process, throughput of 100 jobs per week, and average flow time of 5 weeks, the figures are inconsistent because 100 × 5 = 500 jobs, not 1,000. Such discrepancies may indicate different system boundaries or unreliable data.

5. Bottlenecks and the Theory of Constraints

A bottleneck is a resource whose available capacity is less than, or closest to, the required load. The bottleneck limits the throughput of the system. A constraint is broader: it is any factor that prevents the system from achieving its goal. The constraint may be a machine, labor skill, policy, market, supplier, information system, or quality requirement.

The local utilization of resource i can be estimated as:

where Di is the required load on resource i and Ci is its available effective capacity. The resource with the highest utilization is a candidate bottleneck, but managers must also consider product mix, routing, downtime, setup requirements, and interactions among resources.

The Five Focusing Steps

The Theory of Constraints proposes a continuing improvement cycle:

1. Identify the constraint. Locate the factor that limits system throughput.
2. Exploit the constraint. Obtain more output from existing constraint capacity, for example by reducing downtime, prioritizing quality, ensuring material availability, and moving breaks or inspections away from the constraint where feasible.
3. Subordinate everything else. Align nonconstraint resources and policies with the constraint’s needs. Do not produce merely to keep nonbottlenecks busy.
4. Elevate the constraint. Add capacity through overtime, outsourcing, redesign, equipment, cross-training, or investment.
5. Repeat. Once one constraint is broken, another becomes system-limiting. Avoid allowing old policies or assumptions to become the next constraint.

TOC commonly uses three global measures:

• Throughput: the rate at which the system generates money through sales, or more generally the rate of completed useful output.
• Inventory: money invested in items intended for sale or transformation.
• Operating expense: money spent to convert inventory into throughput.

These measures encourage managers to evaluate system performance rather than local efficiency. A nonbottleneck operating at 100 percent may increase inventory and congestion without increasing throughput.

Bottleneck Protection

Because lost bottleneck time is often lost system capacity, bottlenecks require protection from interruptions. A practical protection system may include:

• a time buffer before the bottleneck;
• a small material buffer after the bottleneck to prevent blocking;
• preventive maintenance scheduled during low-risk periods;
• quality checks before bottleneck processing;
• cross-trained relief staff;
• rules preventing low-value work from consuming bottleneck time; and
• real-time escalation when the buffer approaches a critical threshold.

Buffers should be managed as control signals, not as permission to accumulate unlimited work. Excessive buffers conceal problems, extend lead times, and increase obsolete or damaged inventory.

6. Finite Buffers, Blocking, and Starving

In a line with finite intermediate storage, a station can be unable to work for two different reasons:

• Starving: the station has no input because the upstream station has not supplied material.
• Blocking: the station has completed work but cannot release it because downstream storage or the next station is unavailable.

Both conditions reduce effective capacity. In an unbuffered or tightly coupled line, a short interruption at one station can propagate throughout the system. Finite buffers decouple stations, but they also create trade-offs:

Buffer sizing should reflect demand variability, processing variability, replenishment time, reliability, and the cost of shortage versus inventory. There is no universally optimal buffer size.

A simple conceptual decomposition of a line’s time loss is:

This is the logic of overall equipment effectiveness (OEE), although definitions must be standardized. Availability reflects operating time relative to planned production time; performance reflects speed relative to an ideal rate; and quality reflects good units relative to total units. OEE is useful diagnostically, but it should not be treated as a direct measure of system throughput when resources are not bottlenecks.

7. Flow-Control Policies

Push and Pull

A push system releases work according to a schedule or forecast. It can coordinate complex requirements but may generate excess work-in-process when downstream capacity is unavailable. A pull system authorizes replenishment in response to downstream consumption. Kanban is a common pull-control mechanism.

For a basic kanban system, the number of cards or containers may be estimated as:

where n is the number of kanban containers, D is demand per unit time, L is replenishment lead time, α is a safety factor expressed as a proportion, C is container quantity, and ⌈x⌉ means round upward to the next whole container. The formula assumes relatively stable demand, a defined replenishment lead time, and a consistent container quantity. If demand or lead time is highly variable, safety factors must be evidence-based and reviewed.

Release Control and CONWIP

CONWIP—constant work-in-process—controls the number of jobs released into a line. A job enters only when another job leaves, maintaining a system-wide WIP limit. This approach is useful where product routes vary and conventional kanban cards are difficult to define. The WIP limit should be linked to the desired flow time and demonstrated throughput using Little’s Law.

Sequencing and Priority Rules

Common rules include first-come-first-served, shortest processing time, earliest due date, critical ratio, and bottleneck-first priority. Each rule optimizes different objectives. For example, shortest processing time often reduces average flow time but may disadvantage long jobs; earliest due date supports due-date performance but can increase setup costs. No sequencing rule dominates in every operating environment.

8. Integrated Worked Application

Consider a three-stage assembly line. Stage A has capacity of 40 units per hour, Stage B 30 units per hour, and Stage C 36 units per hour. Customer demand is 32 units per hour.

The nominal utilization levels are:

u_A = 32/40 = 0.80

u_B = 32/30 = 1.067

u_C = 32/36 ≈ 0.889

Stage B is the bottleneck because its capacity is below demand. The maximum sustainable throughput cannot exceed 30 units per hour unless Stage B is improved or demand is redirected. At 30 units per hour, Stage A has 10 units per hour of nominal excess capacity and Stage C has 6 units per hour. Producing 40 units per hour at Stage A would not increase throughput; it would increase WIP before Stage B.

Suppose management can reduce Stage B setup losses and increase capacity from 30 to 34 units per hour. The new bottleneck becomes Stage C, whose capacity is 36 units per hour, but at demand of 32 units per hour it operates at approximately 88.9 percent utilization. The system can now meet demand, but variability may still create queues. Management should therefore protect Stage C, monitor its buffer, and evaluate whether the improvement at Stage B merely transfers the constraint.

9. Illustrative Case: Emergency Department Flow

Background

The following is a fictional illustrative case based on common healthcare operations challenges. A regional emergency department receives an average of 42 patients per hour during peak periods. Registration, triage, diagnostic testing, physician assessment, and discharge or admission are connected through finite waiting areas.

Problem

Patients experience long waits despite high staff utilization. Managers initially propose adding more registration clerks because the registration area appears congested. However, the principal delay occurs after diagnostic imaging, where limited radiologist availability and report variability create a growing queue. The emergency department also releases patients from triage in large waves, increasing arrival variability at imaging.

Analysis

A process observation identifies three problems:

1. Imaging is the effective bottleneck during peak periods.
2. Batch release from triage increases the coefficient of variation of arrivals.
3. Incomplete requests cause rework and consume scarce imaging capacity.

Little’s Law shows that 84 patients in the diagnostic pathway at a throughput of 28 patients per hour imply an average pathway time of 3 hours. Increasing the number of patients in the pathway without increasing throughput would worsen waiting time.

Approach and Solution

The department applies the TOC sequence. It first exploits imaging capacity by standardizing requests, reducing avoidable repeat scans, assigning a rapid escalation channel for equipment faults, and scheduling breaks to preserve peak coverage. It then subordinates upstream release to imaging capacity through a controlled pull signal. A small buffer of ready-to-process patients is maintained, but the number of patients awaiting incomplete information is reduced.

Next, the department elevates the constraint by using validated remote reporting during peak periods and cross-training selected physicians for preliminary interpretation where legally and clinically appropriate. A dashboard tracks imaging queue length, report age, equipment availability, and the rate of patients ready for imaging.

Outcome

The intervention increases effective imaging throughput from 28 to 34 patients per hour and reduces average WIP in the diagnostic pathway. Registration remains busy, but adding registration staff is deferred because it does not constrain total throughput. The department continues monitoring because the physician assessment area may become the next constraint.

Learning Points

The case demonstrates that visible congestion is not necessarily the system bottleneck. Flow improvement requires system-level diagnosis, controlled release, protection of the constraint, and attention to quality and safety. In healthcare, throughput improvements must never override clinical judgment, patient privacy, informed consent, or regulatory requirements.

10. Measurement, Technology, and AI

Effective flow control depends on reliable operational data. Useful measures include throughput, WIP, flow time, queue length, utilization, starving time, blocking time, first-pass yield, downtime, setup time, schedule adherence, and on-time completion. Measurements should use explicit boundaries and consistent definitions.

Digital manufacturing execution systems, warehouse systems, hospital information systems, and workflow platforms can provide near-real-time signals. Process mining can reconstruct actual routes from event logs and identify rework loops or hidden queues. Simulation can test buffer sizes, staffing plans, and priority rules before implementation.

AI may assist with demand forecasting, anomaly detection, predictive maintenance, dynamic scheduling, and queue-risk alerts. Its use requires caution. A forecast is not a capacity guarantee, and an optimization model may reinforce historical priorities that disadvantage particular customers or patient groups. Governance should address data quality, explainability, cybersecurity, privacy, human override, and accountability for decisions. AI should support, not replace, operational expertise and safety controls.

11. Challenges and Limitations

Queuing and bottleneck models are powerful abstractions, but their assumptions can fail in practice.

• Arrival rates may be nonstationary, especially during promotions, emergencies, or seasonal peaks.
• Service times may be correlated because complex jobs arrive together or workers learn over time.
• Resources may be shared across products, creating sequence-dependent setups.
• Priority rules may violate the assumptions of first-come-first-served models.
• Buffers may have finite capacity, causing loss, blocking, or reneging.
• Bottlenecks can shift by product mix, time of day, or maintenance state.
• Average rates can conceal extreme variability and poor tail performance.
• Local utilization improvements can increase total flow time.

Managers should therefore combine analytical models with direct observation, time-series analysis, simulation, sensitivity analysis, and frontline knowledge. Recommendations should be tested under realistic variability rather than only average conditions.

12. Best-Practice Framework

A disciplined flow-improvement program can follow this sequence:

1. Define the system boundary, customer, output, and performance objective.
2. Establish reliable measures of arrival rate, throughput, WIP, flow time, variability, and effective capacity.
3. Map the physical and information flow, including rework and waiting.
4. Identify the current constraint using required load, practical capacity, and observed starvation or blocking.
5. Protect and exploit the constraint before purchasing capacity.
6. Control releases using pull, CONWIP, or an equivalent WIP policy.
7. Reduce variability through standard work, quality at source, maintenance, and demand leveling.
8. Use buffers deliberately and monitor their status as control signals.
9. Elevate the constraint only after lower-cost interventions are assessed.
10. Recalculate the system after every major change because the bottleneck may move.
11. Govern technology and AI with transparent definitions, human accountability, and data controls.
12. Review both average performance and reliability measures, including high-percentile waiting times.

Chapter Summary

Variability converts nominal capacity into waiting, idle time, and unstable flow. Queuing theory shows that utilization approaching 100 percent can produce disproportionately large delays, particularly when arrivals or service times are variable. The M/M/1 formulas provide useful special-case results, while the VUT relationship offers a practical approximation for understanding the combined influence of utilization and variability.

Little’s Law, L = λW, is a general steady-state relationship that connects WIP, throughput, and flow time. It is an essential diagnostic tool, provided that units and system boundaries are consistent. Bottlenecks determine system throughput, and the Theory of Constraints directs managers to identify, exploit, subordinate, elevate, and then reassess the constraint. Finite buffers protect against disruption but also create blocking, starving, inventory, and lead-time trade-offs.

Effective flow control combines pull or WIP-release policies, bottleneck protection, variability reduction, disciplined sequencing, and reliable measurement. Technology, simulation, process mining, and AI can strengthen decisions, but they require sound data and responsible governance. The ultimate objective is not maximum local utilization; it is reliable system throughput with acceptable flow time, quality, cost, and risk.

Key Terms

Variability: Random or systematic fluctuation in arrivals, processing, resources, or routing.

Queue: Entities waiting for service or movement through a process.

Utilization: The proportion of available capacity used by required or actual load.

Arrival rate: Average number of entities entering a system per unit of time.

Service rate: Average number of entities a server can process per unit of time.

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

Coefficient of variation: Standard deviation divided by the mean.

VUT relationship: An approximation linking waiting to variability, utilization, and mean processing time.

Little’s Law: The steady-state relationship L = λW.

Work-in-process: Entities released into a process but not yet completed.

Bottleneck: A resource that limits or most nearly limits system throughput.

Constraint: Any factor that prevents a system from achieving its goal.

Theory of Constraints: A continuous-improvement approach centered on system constraints.

Starving: A condition in which a resource lacks input work.

Blocking: A condition in which a completed resource cannot release its output.

Finite buffer: Intermediate storage with a maximum capacity.

Pull system: A system in which downstream consumption authorizes replenishment.

CONWIP: A control method that maintains a specified amount of work in process.

Throughput: The rate of useful completed output.

Effective capacity: Practical output after losses such as downtime, speed loss, and defects.

Review Questions

Short Answer Questions

1. Why can a system operating at 95 percent utilization experience substantial waiting?
2. State Little’s Law and define each variable.
3. Distinguish starving from blocking in a production line.
4. What is the purpose of subordinating nonbottleneck resources in TOC?
5. Why must system boundaries be defined when applying Little’s Law?

Long Answer Questions

1. Explain how arrival variability and service-time variability affect queue performance. Relate your answer to the VUT relationship.
2. Compare push, pull, and CONWIP flow-control systems, including their advantages and limitations.
3. Describe the five focusing steps of the Theory of Constraints and illustrate each with an operations example.
4. Discuss the trade-offs involved in choosing finite-buffer sizes in a multistage production line.
5. Critically evaluate the use of utilization as a performance measure for nonbottleneck resources.

Analytical / Application Questions

1. A service center receives 18 jobs per hour. Each job requires an average of 2.5 minutes of service from one technician. Assuming an M/M/1 model, calculate utilization, mean time in the system, and mean number in the system. Interpret the results.
2. A process completes 240 units per day and contains an average of 1,200 units of WIP. Management wants a four-day average flow time without changing throughput. Calculate the required WIP reduction and explain two operational actions that could achieve it.
3. A line has three stations with capacities of 55, 48, and 60 units per hour, while demand is 50 units per hour. Identify the bottleneck, calculate each station’s utilization, and recommend a TOC-based improvement sequence.

Multiple Choice Questions with Answers

Question 1

In a stable single-server queue, which condition is necessary for long-run stability?

A. λ > μ

B. λ = μ

C. λ < μ

D. ρ > 1

Correct Answer: C. λ < μ

Question 2

For an M/M/1 queue with λ = 6 jobs/hour and μ = 7.5 jobs/hour, what is utilization?

A. 0.60

B. 0.80

C. 1.25

D. 13.5

Correct Answer: B. 0.80

Question 3

Which relationship connects average WIP, throughput, and average flow time?

A. OEE = Availability × Performance × Quality

B. L = λW

C. ρ = μ/λ

D. n = D + L + C

Correct Answer: B. L = λW

Question 4

A station is starved when it:

A. Cannot release completed work

B. Has no input work available

C. Operates above its design speed

D. Has excessive finished-goods inventory

Correct Answer: B. Has no input work available

Question 5

Which TOC step follows identification of the constraint?

A. Elevate the constraint immediately

B. Subordinate the constraint to nonbottlenecks

C. Exploit the constraint

D. Remove all buffers

Correct Answer: C. Exploit the constraint

Question 6

In the VUT relationship, which factor directly represents service-time variability?

A. Ca

B. Cs

C. u

D. te only

Correct Answer: B. Cs

Question 7

A process has throughput of 100 jobs per day and average WIP of 500 jobs. By Little’s Law, average flow time is:

A. 0.2 days

B. 5 days

C. 400 days

D. 50,000 days

Correct Answer: B. 5 days

Question 8

If a nonbottleneck produces faster than the bottleneck can consume, the most likely immediate result is:

A. Higher system throughput automatically

B. Lower demand variability

C. More WIP before the bottleneck

D. Elimination of the constraint

Correct Answer: C. More WIP before the bottleneck

Question 9

Which statement about finite buffers is most accurate?

A. Larger buffers always minimize total cost

B. Finite buffers can reduce starvation but may increase WIP and lead time

C. Finite buffers eliminate blocking under all conditions

D. Small buffers always maximize throughput

Correct Answer: B. Finite buffers can reduce starvation but may increase WIP and lead time

Question 10

For a line with capacities 40, 30, and 36 units/hour and demand of 32 units/hour, which station is the bottleneck?

A. The 40-unit/hour station

B. The 30-unit/hour station

C. The 36-unit/hour station

D. Demand is always the bottleneck

Correct Answer: B. The 30-unit/hour station