Application of Mathematics in Real-World Usage · Guide 26 · BTT Mathematics Hub
A counter can serve sixty customers an hour and receive only fifty-five, yet queues can still form. A system operating at 99% average utilisation can perform much worse than one at 80% because variability has almost no spare capacity to absorb it. Adding one server can reduce waiting by far more than one server’s proportional share. Queueing Mathematics turns flow, capacity and uncertainty into service-system decisions.
All shops, clinics, call centres, servers and arrival streams below are fictional teaching systems. This is not operational advice for emergency, healthcare, aviation or safety-critical services. Real systems require measured arrival distributions, service distributions, priority rules, abandonment behaviour and constraints that simplified textbook models may omit.
Arrival rate and service rate are flows per time
Let λ be average arrival rate and μ the average service rate of one server. If customers arrive at 12 per hour and one server completes 15 per hour on average, λ=12/h and μ=15/h.
Mean service time is the reciprocal of service rate: 1/15 hour = 4 minutes per customer.
Utilisation compares demand with capacity
For one server, utilisation ρ=λ/μ. With λ=12 and μ=15, ρ=0.8 or 80%.
For c identical parallel servers, nominal capacity is cμ and utilisation is λ/(cμ). If three servers each handle 10 per hour and arrivals average 24 per hour, utilisation is 24/30=80%.
Average capacity above average demand is necessary but not sufficient for zero waiting
Even with λ<μ, random arrivals can bunch together and service times can vary. Temporary queues therefore occur in stochastic systems with spare average capacity.
The condition λ<μ is a stability condition for the standard single-server M/M/1 model, not a promise of no waiting.
An unstable queue grows without a finite steady-state mean
If arrivals average 16 per hour while one server averages 15 per hour, demand exceeds service capacity. Under a persistent model with no abandonment or finite cap, backlog grows.
Temporary operation above capacity can be absorbed by an existing buffer, but long-run average demand above long-run average capacity cannot be cleared indefinitely.
Little’s Law links number, time and throughput
Little’s Law states L=λW for a stable system under broad conditions: average number in system equals throughput rate times average time in system.
If 20 customers per hour pass through and average time in system is 0.25 hour, average number in system is 5.
The units check: customers/hour × hours = customers.
Little’s Law also applies to the waiting line when boundaries match
If average queue length Lq=3 customers and throughput is 12 per hour, average waiting time Wq=3/12 hour=0.25 hour=15 minutes.
The queue-only quantities must use the same boundary. Mixing number-in-system with waiting-before-service gives an inconsistent equation.
The M/M/1 model shows nonlinear waiting near full utilisation
For the idealised M/M/1 queue with Poisson arrivals and exponential service, mean time in system is W=1/(μ−λ).
With μ=15/h and λ=12/h, W=1/3 hour=20 minutes. Mean service time is 4 minutes, so mean waiting before service is 16 minutes.
If λ rises to 14/h with the same μ, W becomes one hour. Arrival demand increased only 16.7% from 12 to14, but mean system time tripled because spare capacity collapsed from 3/h to1/h.
Queue length also rises nonlinearly
For M/M/1, L=ρ/(1−ρ). At ρ=0.8, L=4. At ρ=0.9, L=9. At ρ=0.95, L=19.
Ten percentage points of extra utilisation from 80% to90% more than doubles average number in system in this model.
Waiting probability is not the same as average waiting time
In an M/M/1 system, an arriving customer finds the server busy with probability ρ under the standard stationary assumptions. At 80% utilisation, that probability is 80%.
The customer may still wait only briefly or much longer depending on the queue state. Probability of waiting and duration of waiting are different metrics.
Capacity buffers can be expressed as headroom
A service centre has capacity 100 jobs/hour and average demand 75. Absolute headroom is 25 jobs/hour. Relative headroom against capacity is 25%.
If demand rises to 90, headroom falls to10%. The system may still be stable on average, but variability has less room to be absorbed.
Peak demand and average demand answer different capacity questions
A station averages 60 arrivals/hour across a day but experiences 100/hour during a one-hour peak. Designing only for the daily average can produce severe peak queues.
Capacity planning therefore needs the time scale of the demand statistic: annual average, daily average, peak-hour average or short burst.
Pooling servers can reduce imbalance
Imagine two separate queues with one server each. One queue temporarily has six customers and the other zero. One server is idle while the other queue waits.
A pooled queue feeding both servers can use the idle server for the next customer, reducing imbalance. The exact waiting-time benefit depends on arrival and service distributions.
Priority queues change fairness and delay distribution
Suppose urgent jobs receive priority over routine jobs. The overall throughput may remain unchanged while urgent waiting falls and routine waiting rises.
A single average waiting time can therefore hide large differences between classes. Service policy determines which average is relevant.
Abandonment changes effective throughput
If 100 customers arrive but 8 leave before service, only 92 complete service. Arrival rate, abandonment rate and completed throughput are different flows.
Using arrival rate inside Little’s Law for a boundary that includes only completed customers can be wrong if abandonment is material.
Service-level targets are threshold probabilities
A fictional call centre target might be “80% answered within 30 seconds.” If 1,000 eligible calls are observed and 830 meet the threshold, empirical service level is 83%.
Average waiting time could still be poor if the remaining 170 calls wait extremely long. Threshold service level and mean waiting capture different parts of the distribution.
Percentiles reveal tail waiting
If the 95th percentile wait is 12 minutes, 95% of observed waits are at or below the chosen percentile definition near 12 minutes and 5% exceed it.
A mean of 3 minutes does not prevent a long upper tail. Queue performance often needs both central and tail measures.
Batching changes the service process
A machine processes 20 items every 10 minutes as a batch. Its long-run nominal rate is 120 items/hour, but an item arriving just after a batch starts may wait nearly 10 minutes before processing begins.
Throughput rate alone does not describe latency when work is released in batches.
Finite buffers convert waiting into rejection
Suppose a queue can hold at most ten waiting jobs. An eleventh arrival while all ten places are occupied is blocked or diverted under the model.
The finite buffer prevents unbounded queue length but introduces loss probability. Capacity constraints can move a problem rather than eliminate it.
A simulation can represent systems without closed-form formulas
Generate arrival times and service times from specified distributions, process events in chronological order, and record waiting times and queue lengths. Repeating many simulated days estimates the distribution of outcomes under the assumed model.
Simulation does not rescue unrealistic assumptions. It computes the consequences of the model supplied.
A complete queue report names its boundary and stochastic assumptions
State whether time includes service or only pre-service waiting, whether customers can abandon, whether servers are identical, whether priorities exist and whether arrivals and service are modelled as Poisson/exponential or measured empirically.
Then report utilisation, throughput, mean waiting and a tail measure. One average rarely captures the whole service experience.
Practice: twenty queueing Mathematics questions
- λ=12/h and μ=15/h. Find utilisation.
- Find mean service time.
- Three servers each provide 10/h and λ=24/h. Find utilisation.
- For L=5 and λ=20/h, find W using Little’s Law.
- For Lq=3 and λ=12/h, find Wq.
- For M/M/1 with μ=15 and λ=12, find W.
- Subtract service time to find Wq.
- For μ=15 and λ=14, find W.
- Find L for ρ=0.8.
- Find L for ρ=0.9.
- Capacity 100/h and demand 75/h. Find headroom.
- Demand rises to90/h. Find relative headroom against capacity.
- 100 arrivals with 8 abandon. Find completed-throughput count.
- 830 of1,000 calls meet a threshold. Find service level.
- A batch processes 20 items every10 min. Find nominal hourly throughput.
- If λ=16 and μ=15 in persistent M/M/1 assumptions, is the queue stable?
- If utilisation is 95%, find M/M/1 L.
- If average number in system is 8 and throughput is16/h, find average system time.
- Two separate queues have six and zero customers. What operational inefficiency can pooling reduce?
- Why can waiting rise sharply as utilisation approaches 100%?
Worked answers
- 80%.
- 4 minutes.
- 80%.
- 0.25 h = 15 min.
- 15 min.
- 20 min.
- 16 min.
- 1 hour.
- 4 customers.
- 9 customers.
- 25 jobs/h.
- 10%.
- 92.
- 83%.
- 120 items/h.
- No. Average arrivals exceed service capacity.
- 19.
- 0.5 h = 30 min.
- Idle capacity beside a congested queue.
- Because spare capacity becomes very small, so random arrival and service fluctuations take longer to clear.
Sources and connected applications
For Little’s Law and foundational queueing concepts, see MIT: Little’s Law as Viewed on Its 50th Anniversary. The service systems and numerical cases here are fictional teaching examples.
Continue with Healthcare Screening, Diagnostic Tests, Risk and Rates; Elections, Voting, Apportionment and Ranking; and Insurance, Risk, Premiums, Expected Loss and Pooling. Return to the BTT Mathematics Hub.