2.2 Queueing Theory Models (M/M/1, M/M/s) & Little's Law Applications

2.2 Queueing Theory Models (M/M/1, M/M/s) & Little's Law Applications

Queueing theory is the mathematical study of waiting lines, congestion, and service delays in stochastic operational environments. In industrial and systems engineering, queueing models evaluate manufacturing assembly cells, automated guided vehicle (AGV) dispatching, telecommunication networks, warehouse receiving docks, and healthcare delivery systems to balance the cost of providing service capacity against the cost of customer waiting or work-in-process (WIP) holding.

---

Architecture of Queueing Systems & Kendall's Notation

Every queueing system consists of five physical components: an arrival process from an input population, a waiting line (queue buffer), a service mechanism (servers), a service discipline, and customer departure.

To standardize queueing models, British statistician David G. Kendall developed Kendall's Notation, expressed as:

$A / B / c / K / m / Z$

Where:

• $A$ (Arrival Process): Probability distribution of interarrival times.
  • $M$: Markovian / Exponential (memoryless, Poisson arrival process)
  • $D$: Deterministic (constant interarrival times)
  • $E_k$: Erlang-$k$ distribution
  • $G$: General continuous probability distribution
• $B$ (Service Time Distribution): Probability distribution of service durations.
  • $M$: Exponential service times
  • $D$: Deterministic / constant service times
  • $G$: General service time distribution
• $c$ (Number of Parallel Servers): Integer count ($s$ or $c \ge 1$) of identical, independent service channels operating in parallel.
• $K$ (System Capacity): Maximum permissible number of customers in the system simultaneously (in queue plus in service). When omitted, default is $\infty$.
• $m$ (Calling Population Size): Size of the customer source pool. When omitted, default is $\infty$.
• $Z$ (Queue Discipline): The priority order for service admission. Common disciplines include:
  • FIFO (First-In, First-Out): Standard industrial queue discipline.
  • LIFO (Last-In, First-Out): Found in inventory stackers or bucket elevators.
  • SIRO (Service In Random Order): Used in communication packet routing.
  • PRI (Priority Service): Preemptive or non-preemptive urgent dispatch.

When the last three parameters are omitted (e.g., $M/M/1$ or $M/M/s$), the system implies an infinite queue capacity ($K = \infty$), an infinite calling population ($m = \infty$), and a FIFO discipline.

---

Stochastic Foundations: Poisson Processes & Exponential Distributions

Most classical queueing formulations assume Markovian ($M$) dynamics governed by the memoryless property.

1. The Poisson Arrival Process

Let $\lambda$ denote the mean arrival rate (customers/hour, parts/minute). The probability that exactly $n$ arrivals occur within an observation time window $t$ is given by the Poisson distribution:

$P(N(t) = n) = \frac{(\lambda t)^n e^{-\lambda t}}{n!} \quad \text{for } n = 0, 1, 2, \dots$

• Expected arrivals in interval $t$: $E[N(t)] = \lambda t$
• Variance of arrivals: $\text{Var}(N(t)) = \lambda t$

If the arrival counting process is Poisson, the interarrival times ($T_a$) between successive arrivals are independently and identically distributed (i.i.d.) according to an exponential distribution with parameter $\lambda$:

$f(t) = \lambda e^{-\lambda t} \quad (t \ge 0), \quad F(t) = 1 - e^{-\lambda t}$ $E[T_a] = \frac{1}{\lambda}, \quad \text{Var}(T_a) = \frac{1}{\lambda^2}$

2. Exponential Service Times

Let $\mu$ represent the mean service rate of a single server (parts processed per unit time when continuously busy). The mean service duration per customer is $1/\mu$. The service duration $T_s$ follows:

$f(t) = \mu e^{-\mu t} \quad (t \ge 0), \quad F(t) = 1 - e^{-\mu t}$

3. The Memoryless (Markovian) Property

The exponential distribution is the only continuous distribution possessing the memoryless property:

$P(T > t + s \mid T > s) = P(T > t)$

Physical Meaning: The remaining time until an arrival or service completion does not depend on how much time has already elapsed. In an assembly station with exponential service times, if a job has already been in processing for 10 minutes, the probability that it requires an additional 5 minutes is identical to the probability that a freshly arrived job takes 5 minutes.

---

Universal Conservation Laws: Little's Law & Traffic Intensity

Traffic Intensity (Utilization Factor $\rho$)

The traffic intensity measures the operational load placed on the service facility:

• Single-Server System ($c = 1$): $\rho = \frac{\lambda}{\mu}$
• Multi-Server System ($c = s$ servers): $\rho = \frac{\lambda}{s\mu}$

Stability Criterion (Ergodicity): For any infinite-capacity queue to reach a steady-state equilibrium, the service facility must be capable of processing work faster than the arrival rate:

$\rho < 1 \implies \lambda < s\mu$

If $\rho \ge 1$, the queue length grows toward infinity over time ($L_q \to \infty$), resulting in system instability.

Little's Law

Formulated rigorously by John D. C. Little in 1961, Little's Law is a distribution-free fundamental theorem that relates the average inventory in a system to the throughput rate and the average flow time:

$L = \lambda W$ $L_q = \lambda W_q$

Where:

• $L$: Average number of customers/units in the entire system (queue + service).
• $L_q$: Average number of customers/units waiting in the queue.
• $W$: Average time a customer spends in the entire system (sojourn time).
• $W_q$: Average time a customer spends waiting in the queue.
• $\lambda$: Average arrival rate of entering customers.

Fundamental Additive Decompositions

Total time in the system is the sum of waiting time in queue and processing time in service:

$W = W_q + \frac{1}{\mu}$

Multiplying throughout by $\lambda$ yields the relationship between system and queue inventory:

$L = L_q + \frac{\lambda}{\mu} = L_q + \rho \quad (\text{for } s = 1)$ $L = L_q + \frac{\lambda}{\mu} = L_q + s\rho \quad (\text{for } s \ge 1)$

Little's Law applies universally regardless of arrival distribution, service distribution, or queue discipline, provided the system is stationary and stable.

---

Single-Server Queueing Model ($M/M/1$)

The $M/M/1$ queue represents a single server fed by a Poisson arrival process with rate $\lambda$ and exponential service times with rate $\mu$.

State Probabilities

Using continuous-time Markov chain birth-death transition equations, the steady-state balance equations require: $\lambda P_n = \mu P_{n+1} \implies P_{n+1} = \rho P_n$

Summing all probabilities to unity ($\sum_{n=0}^\infty P_n = 1$) yields the geometric distribution:

$P_0 = 1 - \rho \quad (\text{Probability the system is completely empty / idle})$ $P_n = (1 - \rho)\rho^n \quad (\text{Probability exactly } n \text{ customers are in the system})$

Probability of System Congestion ($N \ge k$)

The probability that the number of entities in the system is at least $k$ is:

$P(N \ge k) = \sum_{n=k}^\infty (1 - \rho)\rho^n = \rho^k$

Performance Metric Closed-Form Equations

1. Average number in system ($L$): $L = \sum_{n=0}^\infty n P_n = \frac{\rho}{1 - \rho} = \frac{\lambda}{\mu - \lambda}$

2. Average number in queue ($L_q$): $L_q = L - \rho = \frac{\lambda}{\mu - \lambda} - \frac{\lambda}{\mu} = \frac{\lambda^2}{\mu(\mu - \lambda)}$

3. Average time in system ($W$): $W = \frac{L}{\lambda} = \frac{1}{\mu - \lambda}$

4. Average time in queue ($W_q$): $W_q = \frac{L_q}{\lambda} = \frac{\lambda}{\mu(\mu - \lambda)}$

The "Hockey-Stick" Congestion Curve

Because the denominator contains $(\mu - \lambda) = \mu(1 - \rho)$, as utilization $\rho \to 1.0$, the waiting time curve escalates non-linearly. At $\rho = 0.50$, $L = 1.0$; at $\rho = 0.80$, $L = 4.0$; at $\rho = 0.95$, $L = 19.0$. Industrial systems designed to operate at near 100% utilization experience extreme queue inflation whenever stochastic variability is present.

---

Multi-Server Queueing Model ($M/M/s$)

The $M/M/s$ model features a single waiting queue feeding $s$ parallel, identical servers, each processing at rate $\mu$. Total service capacity is $s\mu$.

Transition Dynamics

The aggregate service rate depends on the number of entities in the system $n$: $\mu_n = \begin{cases} n\mu & \text{if } 0 \le n < s \\ s\mu & \text{if } n \ge s \end{cases}$

Probability of an Empty System ($P_0$)

$P_0 = \left[ \sum_{n=0}^{s-1} \frac{(\lambda / \mu)^n}{n!} + \frac{(\lambda / \mu)^s}{s! \left(1 - \frac{\lambda}{s\mu}\right)} \right]^{-1}$

Erlang C Formula (Probability of Waiting)

The probability that an arriving customer finds all $s$ servers occupied and must wait in the queue is given by the Erlang C formula ($C(s, \lambda/\mu)$):

$P(\text{wait}) = \frac{\frac{(\lambda / \mu)^s}{s!(1 - \rho)}}{ \sum_{n=0}^{s-1} \frac{(\lambda / \mu)^n}{n!} + \frac{(\lambda / \mu)^s}{s!(1 - \rho)} } = P_0 \cdot \frac{(\lambda / \mu)^s}{s!(1 - \rho)}$

Multi-Server Operational Metrics

1. Average queue length ($L_q$): $L_q = \frac{P_0 (\lambda / \mu)^s \rho}{s!(1 - \rho)^2} = \frac{P(\text{wait}) \cdot \rho}{1 - \rho}$

2. Average time in queue ($W_q$): $W_q = \frac{L_q}{\lambda}$

3. Average time in system ($W$): $W = W_q + \frac{1}{\mu}$

4. Average number in system ($L$): $L = L_q + \frac{\lambda}{\mu} = \lambda W$

---

Engineering Comparative Analysis: Single Fast Server vs. Server Pooling

Consider an industrial inspection facility handling an arrival rate of $\lambda = 16$ assemblies per hour. Management must decide between three design configurations:

• Configuration A (Single Fast Server): One advanced automated cell with $\mu = 20$ parts/hour ($M/M/1$).
• Configuration B (Pooled Parallel Servers): Two standard inspection cells pooled with a shared queue, each operating at $\mu = 10$ parts/hour ($M/M/2$).
• Configuration C (Unpooled Dedicated Servers): Two separate lines, each with its own dedicated queue, splitting arrivals evenly so $\lambda = 8$ parts/hour and $\mu = 10$ parts/hour ($2 \times M/M/1$).

Quantitative Calculation Comparison

Configuration A: $M/M/1$ with $\lambda = 16, \mu = 20$

• Utilization: $\rho = \frac{16}{20} = 0.80$
• $L = \frac{16}{20 - 16} = 4.00$ assemblies
• $L_q = L - \rho = 4.00 - 0.80 = 3.20$ assemblies
• $W = \frac{1}{20 - 16} = 0.25$ hours $= 15.00$ minutes
• $W_q = \frac{3.20}{16} = 0.20$ hours $= 12.00$ minutes

Configuration B: $M/M/2$ with $\lambda = 16, \mu = 10, s = 2$

• Ratio $\frac{\lambda}{\mu} = \frac{16}{10} = 1.60$; Utilization $\rho = \frac{1.60}{2} = 0.80$
• Calculate $P_0$: $P_0 = \left[ \frac{1.6^0}{0!} + \frac{1.6^1}{1!} + \frac{1.6^2}{2!(1 - 0.80)} \right]^{-1} = \left[ 1 + 1.6 + \frac{2.56}{2(0.20)} \right]^{-1} = [2.6 + 6.4]^{-1} = \frac{1}{9} \approx 0.1111$
• Calculate $P(\text{wait})$: $P(\text{wait}) = P_0 \cdot \frac{1.6^2}{2!(1 - 0.80)} = \frac{1}{9} \cdot 6.40 = 0.7111$
• Calculate $L_q$: $L_q = \frac{P(\text{wait}) \cdot \rho}{1 - \rho} = \frac{0.7111 \cdot 0.80}{0.20} = 2.844 \text{ assemblies}$
• Calculate $W_q$: $W_q = \frac{L_q}{\lambda} = \frac{2.844}{16} = 0.1778 \text{ hours} = 10.67 \text{ minutes}$
• Calculate $W$: $W = W_q + \frac{1}{\mu} = 0.1778 + 0.1000 = 0.2778 \text{ hours} = 16.67 \text{ minutes}$
• Calculate $L$: $L = L_q + \frac{\lambda}{\mu} = 2.844 + 1.600 = 4.444 \text{ assemblies}$

Configuration C: Unpooled $2 \times (M/M/1)$ with $\lambda = 8, \mu = 10$

• Utilization: $\rho = \frac{8}{10} = 0.80$
• $L = \frac{8}{10 - 8} = 4.00$ assemblies per line (Total across 2 lines $= 8.00$)
• $L_q = \frac{8^2}{10(10 - 8)} = \frac{64}{20} = 3.20$ assemblies per line (Total across 2 lines $= 6.40$)
• $W = \frac{1}{10 - 8} = 0.50$ hours $= 30.00$ minutes
• $W_q = \frac{8}{10(2)} = 0.40$ hours $= 24.00$ minutes

System Comparison Summary Table

Key Industrial Insights

1. The Pooling Principle: Moving from unpooled queues (Config C) to a shared pooled queue (Config B) with identical total capacity reduces average queue wait time ($W_q$) from 24.0 minutes to 10.67 minutes—a 55.5% reduction in waiting time with zero additional investment in machinery.
2. Fast Single Server Trade-Off: Configuration A achieves the shortest overall system time ($W = 15.0$ minutes) because its service time is half as long (3 min vs 6 min), but Configuration B provides lower queue waiting time ($W_q = 10.67$ min) and provides redundancy (if one server fails, the facility continues operating at 50% capacity, whereas Config A experiences complete line shutdown).