Founded by Agner Krarup Erlang (The Theory of Probabilities and Telephone Conversations, 1909), with David G. Kendall's notation (1953), Leonard Kleinrock's Queueing Systems (1975), and the textbook of Frederick S. Hillier and Gerald J. Lieberman (Introduction to Operations Research, 1967).
Models waiting lines from arrival rates, service rates, and server counts to predict waits, queue lengths, and utilization, showing how waits explode as utilization nears full.
Waits at a support desk explode as the agents get busier
Calls arrive at random and each takes six minutes on average. The curve is the average wait at each level of busyness for the number of agents set below; it stays low through the middle and climbs steeply as the agents near full use. The square is your target wait at today's busyness. Below the curve the desk misses the target; add an agent and watch the busyness fall and the square clear the curve. Illustrative.
At today's callsWithin target: the average wait is shorter than the wait you accept.Over target: callers wait longer than you accept, so add an agent or cut the handling time.
An example
A software company's support desk took about 30 calls an hour at midday, each lasting six minutes on average, with four agents on the phones. Customers were complaining about hold times, but the operations manager's dashboard showed the agents busy only 75 percent of the time, and he did not see the case for a fifth agent who would sit idle even more.
The team's analyst ran the desk through the standard queueing model for random arrivals. At 75 percent busy with four agents, about half of callers waited, and the average wait before an agent answered was a little over three minutes. A fifth agent would bring busyness down to 60 percent and the average wait to under a minute. A sixth would cut it to seconds.
The manager had been reading idle time as waste. It was the slack that absorbed calls bunching up at random. He added a fifth agent for the four midday hours, kept four the rest of the day, and set the desk's target as an average wait of two minutes, measured hourly.
Why it works
Arrivals at a service desk are random, so even when the average load is well under capacity, calls bunch up. During a bunch a line forms, and it takes the desk's spare capacity to clear it. The less spare capacity there is, the longer each line takes to clear and the sooner the next bunch arrives before it has. That is why waits do not grow steadily with busyness: they stay small through the middle range and then climb steeply as busyness approaches full, exploding in the last few points.
The practical result runs against intuition. Running servers near full use, which looks efficient on a utilization report, is what makes the waits long. A modest amount of idle time is the price of short waits, and adding one server to a busy desk usually cuts the wait far more than in proportion.
The thinking behind it
Agner Krarup Erlang Agner Krarup Erlang, a Danish mathematician and engineer at the Copenhagen Telephone Company, founded teletraffic engineering and queueing theory with The Theory of Probabilities and Telephone Conversations (1909), published in Danish in Nyt Tidsskrift for Matematik. The paper showed that random telephone calls follow a Poisson distribution, the starting assumption of the model in this page's figure. The unit of traffic load, the erlang, is named after him; a desk with 30 six-minute calls an hour carries three erlangs.
David G. Kendall David G. Kendall, a British probabilist and statistician, analyzed queues through an imbedded Markov chain in Stochastic Processes Occurring in the Theory of Queues and their Analysis by the Method of the Imbedded Markov Chain (1953), in the Annals of Mathematical Statistics. The paper also proposed the notation that names queueing models by their arrival process, their service distribution, and their number of servers, as in M/M/1 or the M/M/c model behind this page's figure.
Leonard Kleinrock Leonard Kleinrock, a computer scientist, applied queueing theory to computer networks and took part in the development of the ARPANET. His Queueing Systems, Volume 1: Theory (1975), published by Wiley-Interscience, is a standard text on the subject. His work carried the mathematics of telephone lines into computing, where the same question, how much spare capacity keeps waits short, decides the size of servers and network links.
How to run it
- Measure the arrivals. Count arrivals per hour, by hour of the day, since a daily average hides the peaks.
- Measure the service time. Find the average time each arrival takes to serve, and how much it varies.
- Compute busyness. Divide arrivals times service time by the number of servers; above 85 percent, expect long waits.
- Model the wait. Use the standard formula for random arrivals and several servers to get the chance of waiting and the average wait.
- Set a service target. Agree on the wait customers will accept, in minutes or as a share answered within a time.
- Staff to the target by hour. Find the fewest servers that meet the target in each hour, and schedule to that.
- Check against reality. Compare predicted and measured waits each week, and refit if arrivals or service times have shifted.
Where it goes wrong
Busyness is managed as the goal. A desk pushed to 95 percent utilization looks efficient and has waits that drive customers away; spot it when the utilization report improves while complaints rise. Manage to the wait target and accept the idle time it needs.
Averages hide the peaks. Staffing to the daily average arrival rate leaves the midday hours far over capacity. Model each hour separately.
The model's assumptions do not hold. Callers who hang up, calls that return, or service times that vary wildly change the answer. Compare the model's predictions with measured waits, and switch to a simulation when they part ways.
When to use something else
- Theory of Constraints. When the work passes through several stations and the question is which one limits the whole flow.
- Monte Carlo Simulation. When arrivals, service times, or behavior are too irregular for the standard formulas.
- Marginal Analysis. When the question is whether the next server is worth its cost against the shorter waits it buys.
Where it sits in SOLVE eX
Go to the source
- The Theory of Probabilities and Telephone Conversations, 1909
- Stochastic Processes Occurring in the Theory of Queues and their Analysis by the Method of the Imbedded Markov Chain, Annals of Mathematical Statistics 24(3), 1953
- Queueing Systems, Volume 1, 1975
- Introduction to Operations Research, queueing chapters, 1967