Queuing Theory: Models, Metrics & Real-World Applications

TAKEAWAYS

• Queuing theory analyzes customer arrival rates (λ), service rates (μ), and waiting times (W) to optimize system efficiency.
• The M/M/1 model (Markovian arrivals, Markovian service, 1 server) is the simplest but widely used for single-channel systems like bank tellers or eSewa payment counters.
• Little’s Law ($L=\lambda W$) links average queue length ($L$), arrival rate (λ), and waiting time ($W$)—critical for designing call centers or hospital emergency rooms.
• Cost-volume analysis balances service costs vs. waiting costs to find the optimal number of servers (e.g., Ncell customer service agents).
• Priority queues (e.g., VIP lanes at Kathmandu airport) and bulk-service systems (e.g., Daraz’s batch order processing) require specialized models.
• Simulation (e.g., Pathao driver dispatch systems) is used when analytical models are too complex.

1. What Is Queuing Theory?

Queuing theory is the mathematical study of waiting lines—how customers, data packets, or jobs arrive, wait, and are served by a system. It helps businesses:

• Reduce customer frustration (e.g., long waits at NTC offices).
• Lower operating costs (e.g., hiring too many vs. too few bank tellers).
• Improve system reliability (e.g., WhatsApp message delivery delays).

Key Terms

2. Queuing Models: The Big 3

Models are classified by arrival pattern (M/G), service pattern (M/G), and number of servers (c).

A. M/M/1: Single-Server Model (Most Common)

• M/M/1: Markovian arrivals, Markovian service, 1 server.
• Assumptions:
  • Arrivals follow a Poisson process (random, independent).
  • Service times are exponentially distributed (e.g., bank teller transactions).
  • Infinite queue capacity (no limit on waiting).
• Key Formulas:
  • Utilization: $\rho =\lambda /\mu$ (must be $<1$ to avoid infinite queues).
  • Average queue length: $L_q={\rho }^2/(1-\rho )$.
  • Average waiting time: $W_q=\rho /(\mu (1-\rho ))$.

WORKED EXAMPLE: NTC Customer Service NTC receives 12 calls/hour ($\lambda =12$). A single agent handles 15 calls/hour ($\mu =15$).

1. Calculate $\rho =12/15=0.8$.
2. Average waiting time: $W_q=0.8/(15(1-0.8))=0.333$ hours = 20 minutes.
3. Problem: If $\lambda$ increases to 16, $\rho =1.066>1$ → queue explodes! Solution: Hire a second agent.

MERMAID DIAGRAM:

flowchart TD
    A["Arrivals (λ=12/hour)"] --> B["Queue"]
    B --> C["Server (μ=15/hour)"]
    C --> D["Departures"]
    D -->|"Feedback"| A

B. M/M/c: Multi-Server Model (e.g., Bank ATMs, Airports)

• c servers (e.g., 3 ATMs at a bank).
• Key Formulas (use Erlang C formula for $W_q$):
  • $L_q=\frac{P_0(\lambda /\mu {)}^c\rho }{c!(1-\rho {)}^2}$
  • $W_q=L_q/\lambda$ Where $P_0$ is the probability of an empty system.

WORKED EXAMPLE: Nabil Bank ATM

• Arrivals: 30 customers/hour ($\lambda =30$).
• Service rate per ATM: 10 customers/hour ($\mu =10$).
• Number of ATMs (c): 3.

1. $\rho =\lambda /(c\mu )=30/30=1.0$ → Critical! Need more ATMs or faster service.
2. If $c=4$: $\rho =30/40=0.75$.
   • $W_q\approx 0.125$ hours = 7.5 minutes.

C. M/G/1: General Service Times (e.g., Daraz Delivery)

• Service times are not exponential (e.g., delivery trucks take variable times).
• Key Formula (Pollaczek-Khinchine):
  • $W_q=\frac{\lambda E[S^2]}{2(1-\rho )}+\frac{1}{\mu }$ Where $E[S^2]$ = variance of service times.

WORKED EXAMPLE: Pathao Driver Dispatch

• Arrivals: 24 orders/hour ($\lambda =24$).
• Average service time: 15 minutes ($1/\mu =0.25$ hours).
• Variance: $E[S^2]=0.04$ (service times vary).

1. $\rho =24\times 0.25=0.6$.
2. $W_q=\frac{24\times 0.04}{2(1-0.6)}+0.25=0.6+0.25=0.85$ hours = 51 minutes. Problem: Too long! Solutions:
   • Add more drivers ($c=2$).
   • Optimize routes (reduce $E[S^2]$).

3. Queue Discipline: Who Gets Served Next?

MERMAID DIAGRAM: Priority Queue

4. Cost-Volume Analysis: How Many Servers Are Optimal?

Businesses must balance:

• Service cost (hiring more staff).
• Waiting cost (customer dissatisfaction, lost sales).

Formula: $$\text{Total Cost}=C_s\times c+C_w\times L_q$$ Where:

• $C_s$ = Cost per server (e.g., ₹50,000/month per teller).
• $C_w$ = Cost per customer waiting (e.g., ₹100/hour lost sales).

WORKED EXAMPLE: ABC Bank Drive-Thru

• Given:
  • $\lambda =15$ customers/hour.
  • Service rate $\mu =20$ customers/hour.
  • $C_s=₹50,000$/month per teller.
  • $C_w=₹100$/hour per customer.
• For 1 teller ($c=1$):
  • $\rho =15/20=0.75$.
  • $L_q=(0.75{)}^2/(1-0.75)=3$ customers.
  • Monthly waiting cost: $3\times 24\times 30\times ₹100=₹216,000$.
  • Total cost: $₹50,000+₹216,000=₹266,000$.
• For 2 tellers ($c=2$):
  • $\rho =15/40=0.375$.
  • $L_q\approx 0.21$ customers.
  • Monthly waiting cost: $₹15,120$.
  • Total cost: $₹100,000+₹15,120=₹115,120$. Optimal choice: 2 tellers (lower total cost).

5. Real-World Applications in Nepal & Globally

A. Nepali Examples

1. NTC Call Centers

   • Model: M/M/c (multiple agents).
   • Problem: High $\lambda$ during peak hours (e.g., 50 calls/hour).
   • Solution: Use Erlang C to calculate optimal agents (e.g., $c=5$ for $\rho <0.8$).

2. Khalti/E-Sewa Payment Queues

   • Model: M/G/1 (variable service times for transactions).
   • Issue: Long waits during festivals (e.g., Dashain).
   • Fix: Add priority lanes for high-value transactions.

3. Daraz Delivery Hubs

   • Model: Multi-stage queuing (orders → sorting → dispatch).
   • Optimization: Use SPT to dispatch shortest-delivery orders first.

B. Global Examples

1. Amazon Warehouses

   • Model: Bulk-service queues (batch picking).
   • Tech: RFID tags reduce service time variance ($E[S^2]$).

2. Airport Security (e.g., Heathrow)

   • Model: M/M/c with priority lanes (families, business class).
   • Cost Analysis: Balances scanner costs vs. passenger delays.

3. Netflix Streaming

   • Model: M/G/1 for data packets (variable service times).
   • Solution: Buffering acts as a queue to smooth delays.

CASE STUDY: Toyota’s Just-in-Time (JIT) System

• Problem: Long queues at assembly lines caused delays.
• Solution: Applied queuing theory to:
  • Optimize arrival rates of parts (Poisson process).
  • Use kanban systems (visual signals) to limit $L_q$.
• Result: Reduced waiting time by 40%, cutting costs.

6. Advanced Topics

A. Bulk-Service Queues (e.g., Daraz Batch Orders)

• Customers arrive individually, but the server processes them in batches.
• Example: A delivery driver picks up 5 orders at once.
• Formula: $$W_q=\frac{\lambda E[S^2]}{2(1-\rho )}+\frac{E[S]}{2}$$

B. Priority Queues (e.g., Hospitals, Airlines)

• Preemptive priority: High-priority customers interrupt ongoing service.
• Non-preemptive: Serve current customer first, then priority.
• Example: A heart attack patient jumps ahead of a routine check-up.

C. Simulation (When Analytical Models Fail)

• Tools: Arena, AnyLogic, Excel.
• Example: Pathao uses Monte Carlo simulation to model driver availability and traffic delays.

Exam Tip: How to Score Full Marks

1. Always define key terms (e.g., "Queuing theory is the study of waiting lines...").
2. Draw diagrams for models (M/M/1, M/M/c) and label $\lambda$, $\mu$, $\rho$, $L_q$.
3. Show calculations step-by-step (e.g., $\rho =\lambda /\mu$, then $W_q=\rho /(\mu (1-\rho ))$).
4. Relate to Nepal:
   • NTC call centers → M/M/c.
   • Daraz deliveries → M/G/1 or bulk-service.
   • Bank ATMs → Cost-volume analysis.
5. Compare models in a table (e.g., M/M/1 vs. M/M/c assumptions).
6. Discuss trade-offs:
   • "More servers reduce $W_q$ but increase costs."
7. Use real numbers from examples (e.g., "If $\lambda =20$ and $\mu =15$, $\rho =1.33$ → unstable queue").

Common Mistakes to Avoid:

• Forgetting $\rho <1$ is required for stability.
• Mixing up $L_q$ (queue length) and $L$ (total in system).
• Ignoring units (e.g., $\lambda$ in customers/hour, not just numbers).