Queuing Theory: Models, Metrics, and Real-World Applications

TAKEAWAYS:

• Queuing theory models customer flow (arrivals, service times) to optimize wait times and resource use in systems like banks, hospitals, and ride-hailing apps.
• Kendall’s notation (e.g., M/M/1) encodes arrival/service distributions and server channels—critical for matching models to real problems.
• Little’s Law ($L=\lambda W$) links average queue length ($L$), arrival rate ($\lambda$), and wait time ($W$)—use it to solve for any unknown.
• Single-channel vs. multi-channel systems differ in stability (e.g., M/M/1 can have infinite queues, while M/M/c cannot).
• Cost-volume analysis balances service costs vs. wait-time costs to find the optimal number of servers (e.g., how many eSewa counters to open).
• Exam focus: Memorize formulas for $L_q$, $W_q$, $L$, and $W$; recognize when to use M/M/1, M/M/c, or M/G/1; and interpret real-world scenarios (e.g., NTC call centers, Daraz delivery queues).

1. What is a Queue? Core Definitions

A queue is a line of entities (customers, data packets, vehicles) waiting for service from one or more servers. Queuing theory studies how to design systems to minimize waiting time and cost while maximizing throughput.

Key Terms Visualized

Real-World Analogy:

• eSewa counters: Customers (arrivals) wait for service (bill payment) at a single server (counter). If the counter is busy, the queue grows.
• Pathao drivers: Riders (arrivals) wait for drivers (servers) to accept their requests. The "service time" is the time from request to pickup.

2. Arrival and Service Patterns

Queuing theory assumes arrivals and service times follow probability distributions. The most common are:

Worked Example 1: Arrival Rate Calculation Problem: In a clinic, 96 patients arrive in 24 hours. What is the arrival rate ($\lambda$) in patients per minute? Visual:

3. Queue Disciplines: How Customers Are Served

The queue discipline defines the order in which customers are served. Common types:

flowchart TD
    A["Queue Discipline"] --> B["FIFO (First-In-First-Out)"]
    A --> C["LIFO (Last-In-First-Out)"]
    A --> D["SIRO (Service In Random Order)"]
    A --> E["Priority"]
    A --> F["Shortest Processing Time (SPT)"]
    B -->|"Example"| G["eSewa counters"]
    E -->|"Example"| H["NTC VIP calls"]
    F -->|"Example"| I["Pathao driver assignment"]

Comparison Table:

4. Single-Channel Queuing Model (M/M/1)

The simplest model: one server, arrivals follow a Poisson process (rate $\lambda$), service times are exponentially distributed (rate $\mu$).

Key Metrics

Worked Example 2: Clinic Queue Analysis Problem: A clinic has $\lambda =0.00667$ patients/minute and $\mu =0.1$ patients/minute (service time = 10 minutes). Find:

1. Utilization ($\rho$),
2. Average queue length ($L_q$),
3. Average wait time ($W_q$).

Solution:

1. $\rho =\lambda /\mu =0.00667/0.1=0.0667$ (stable, since $\rho <1$).
2. $L_q={\rho }^2/(1-\rho )=(0.0667{)}^2/(1-0.0667)\approx 0.0046$ patients.
3. $W_q=L_q/\lambda =0.0046/0.00667\approx 0.69$ minutes (41 seconds).

Visual:

Real-World Tie-In:

• NTC Call Center: If calls arrive at $\lambda =12$ calls/hour and each call takes $\mu =6$ calls/hour (10 minutes), $\rho =2$ (unstable! NTC needs more agents).
• eSewa Counter: If 10 customers/hour arrive and the agent processes 15/hour, $\rho =0.667$ (efficient, but occasional waits).

5. Multi-Channel Queuing Model (M/M/c)

Multiple servers (e.g., $c=3$ eSewa counters) reduce wait times. Key formulas:

Worked Example 3: Optimal Number of eSewa Counters Problem: eSewa has $\lambda =20$ customers/hour. Each agent serves $\mu =10$ customers/hour. Find:

1. Minimum $c$ for $\rho <0.8$,
2. $L_q$ and $W_q$ for $c=3$.

Solution:

1. For $c=2$: $\rho =20/(2\times 10)=1.0$ (unstable). For $c=3$: $\rho =20/30=0.667<0.8$ (stable).
2. Using Erlang C:
   • $P_0=[1+0.667+(0.667{)}^2/2{]}^{-1}\approx 0.385$.
   • $L_q\approx 0.385\times (0.667{)}^3\times 30^2/[(30-20{)}^2\times 2]\approx 0.55$ customers.
   • $W_q=0.55/(20/60)=1.65$ minutes.

Visual:

Cost-Volume Analysis: eSewa must balance:

• Cost of adding a counter: Rs. 50,000/month per agent.
• Cost of waiting: Rs. 1,000/hour per customer (lost business). For $c=3$, $W_q=1.65$ minutes → Rs. 27.5/hour/customer. Total wait cost = $20\times 27.5=Rs.550/hour$. If adding a 4th counter reduces $W_q$ to 0.5 minutes (Rs. 8.33/hour/customer), the savings may justify the cost.

6. Special Cases and Extensions

Worked Example 4: Pathao Driver Queue (M/M/1/K) Problem: A Pathao driver accepts requests with $\lambda =0.5$ requests/hour and serves $\mu =1$ request/hour. Max passengers $K=3$. Find:

1. Probability the driver is idle ($P_0$),
2. Probability a rider is rejected (queue full).

Solution:

1. $P_0=[1+0.5+(0.5{)}^2/2!+(0.5{)}^3/3!{]}^{-1}\approx 0.45$.
2. Probability of 3 riders: $P_3=(0.5{)}^3/3!\times P_0\approx 0.025$. Rejection rate = 2.5%.

Visual:

7. Queuing Theory in Nepal: Real-World Applications

mindmap
  root((Queuing Theory in Nepal))
    eSewa
      "Single-server model (M/M/1) for counters"
      "Multi-server (M/M/c) for peak hours"
    NTC
      "Call center: M/M/c with priority queues"
      "Unstable if ρ > 1 (too many calls)"
    Pathao
      "Driver assignment: SPT discipline"
      "Queue length depends on rider density"
    Daraz
      "Order processing: M/G/1 (variable delivery times)"
      "Warehouse queues: M/M/c"
    Banks
      "ATM queues: M/M/1 with finite capacity"
      "Teller queues: Priority for VIPs"
    NEPSE
      "Trading system: M/M/c with time limits"
      "Queue discipline: FIFO for orders"

Case Study: NTC Call Center Optimization

• Current: 50 agents, $\lambda =60$ calls/hour, $\mu =12$ calls/hour/agent. $\rho =60/(50\times 12)=0.1$ (underutilized!).
• Problem: Long wait times due to inefficient routing.
• Solution:
  1. Reduce $\lambda$ with IVR (auto-attendant).
  2. Add priority queues for VIP customers.
  3. Use Erlang C to optimize agent count.

Visual:

8. Exam Tip: How to Score Full Marks

1. Memorize Kendall’s Notation:

   • Always state the model (e.g., "This is an M/M/1 queue") before solving.
   • Example: "For the clinic, arrivals are Poisson ($\lambda =0.00667$) and service is exponential ($\mu =0.1$), so it’s M/M/1."

2. Show All Steps:

   • Write down every formula and substitution. Examiners reward clarity.
   • Example:

     Given: λ = 20/hour, μ = 10/hour, c = 3.
     Step 1: ρ = λ/(cμ) = 20/30 = 0.667.
     Step 2: Use Erlang C to find L_q.
     

3. Interpret Results:

   • Don’t just compute $L_q=0.55$; explain: "This means the average queue length is 0.55 customers, so eSewa can reduce counters from 3 to 2 without significant wait increases."

4. Real-World Links:

   • Tie problems to Nepalese services (e.g., "Like NTC call centers, this problem has unstable queues when ρ > 1").
   • Use local examples in explanations (e.g., "Pathao drivers use SPT to minimize wait times").

5. Graphical Answers:

   • Draw queue diagrams (even roughly) to show system structure.
   • Plot cost vs. number of servers to justify optimal $c$.

6. Common Pitfalls:

   • Forgetting stability: Always check $\rho <1$ for M/M/1.
   • Mixing L and L_q: $L$ = in system; $L_q$ = in queue only.
   • Ignoring units: $\lambda$ is per hour/minute; ensure consistency.

9. Practice Problems (Exam-Style)

1. Ncell Hotline: Calls arrive at $\lambda =15$ calls/hour. Each call takes $\mu =10$ calls/hour.

   • Is the system stable? If not, how many agents are needed for $\rho =0.8$?
   • Find $W_q$ for the stable case.

2. Kathmandu Traffic Light: Cars arrive at $\lambda =0.5$ cars/second. The light turns green every 30 seconds (service time = 20 seconds).

   • Model this as M/D/1. Find $L$ and $W$.
   • Suggest improvements using queuing theory.

3. Daraz Warehouse: Orders arrive at $\lambda =30$ orders/hour. Packing takes $\mu =20$ orders/hour.

   • If Daraz adds 2 more packers ($c=3$), what is the new $W_q$?
   • Calculate the cost savings if wait-time cost is Rs. 500/order/hour.

10. Summary Table for Quick Revision

11. Final Visual: Queuing Theory in Action