How busy is too busy? Wait time vs utilization, with Kingman's formula

Average wait grows with utilization ÷ (1 − utilization), multiplied by how variable arrivals and service are. At 80% busy that factor is 4, at 90% it is 9 and at 95% it is 19. Enter your numbers for Kingman's estimate, then check it against a seeded simulation.

80% busy4.0×wait ÷ service time, random arrivals and service
90% busy9.0×more than double the wait at 80%
95% busy19.0×about double 90%

One busy resource

cv is the standard deviation divided by the mean. Random (Poisson) arrivals have cv 1; so do exponential service times. Starting values are an example. One server, unlimited line.

Result

Average wait before service (Kingman)28.3 min 5.67 × service time
Utilization factor u ÷ (1 − u)5.67
Variability factor1 (ca² + cs²) ÷ 2
0×20×40×50%60%70%80%90%utilization
Average wait in multiples of one service time (Kingman). The red dot is your utilization. Past 90% the curve is nearly vertical.
What-if changes and the Kingman wait that results
What changesKingman waitVersus now
Utilization 10 points lower15 min-47%
Service variability cut in half17.7 min-38%
Both together9.4 min-67%
Simulate this in Simflowly

The VUT equation

Wq ≈ V × U × T V = (ca² + cs²) ÷ 2 variability: arrivals and service U = ρ ÷ (1 − ρ) utilization T = average service time

This is Kingman's approximation for one server, usually called the VUT equation. ca and cs are the coefficients of variation (standard deviation ÷ mean) of the time between arrivals and of the service time. Random arrivals and exponential service both have a cv of 1, so V = 1 and the equation reduces to the exact M/M/1 result.

Curve at 70 to 99%

Utilization factor u over one minus u at common utilization levels
Utilization50%70%80%85%90%95%99%
Wait ÷ service time (V = 1)1×2.3×4.0×5.7×9.0×19.0×99.0×

With random arrivals and service, a customer waits on average about as long as one service time at 50% busy, four at 80% and nineteen at 95%. Multiply by V for other variability.

Variability matters as much as load

Same server, same 80% busy, random arrivals, only the spread of service time changes. V runs from 0.5 for constant service to 1.625 for a cv of 1.5, and the wait follows it. The Erlang C vs simulation guide shows the same effect with three servers: average wait of 0.66 minutes with constant service against 2.04 with a long-tailed one.

Simulation check: where Kingman holds and where it does not

One server, average service time of 1, wait measured in service times. Simulation: seed 7, 30 replications of 200 hours after a 10-hour warm-up, with the interval shown. The right-hand column is Kingman divided by the simulation.

Kingman against simulated wait for nine utilization and variability settings
BusyArrivalsServiceKingmanSimulation (95% interval)Kingman vs sim
80%randomrandom4.003.95 (3.80 to 4.10)+1%
90%randomrandom9.008.99 (8.28 to 9.69)+0%
95%randomrandom19.0018.43 (16.01 to 20.86)+3%
80%randomconstant2.002.01 (1.94 to 2.07)0%
80%randomcv 0.52.502.52 (2.44 to 2.61)-1%
80%randomcv 1.56.506.74 (6.25 to 7.23)-4%
90%randomcv 1.514.6315.26 (13.35 to 17.17)-4%
80%cv 0.5cv 0.51.000.83 (0.81 to 0.86)+20%
80%constantcv 0.50.500.35 (0.34 to 0.36)+42%

Honest reading: with random arrivals Kingman is within about 5% of the simulation at every setting here, including 95% busy. When arrivals are more regular than random (cv 0.5 or clockwork) it runs 20% to 42% high in this table, so for scheduled arrivals treat it as an upper bound and simulate.

Limits

  • One server and an unlimited line. Several servers are in the queue calculator (random service only) or a simulation.
  • Steady state. A queue that is still filling after a busy hour is not in the average.
  • Wait is the average. The 95th percentile is several times longer; the simulation reports it.
  • The simulation uses a lognormal shape for any cv other than 0 and 1. Real service times may have other shapes.

Questions

Is 85% the right utilization target?

It is a rule of thumb and no more. The right target depends on how variable arrivals and service are and how many servers share the load: more servers can run hotter for the same wait, and more variable work needs more slack. The calculator shows the wait at your own variability.

Is Kingman's formula exact?

It is exact for one server with random arrivals and exponential service (M/M/1) and an approximation otherwise. In the table below it is within 4% of the simulation whenever arrivals are random, and up to 42% too high when arrivals are more regular than random.

What about more than one server?

Approximations for several servers exist (Allen and Cunneen is a well-known one), but this page does not use them: for several servers use the queue calculator for random service, or simulate. The Simflowly check on this page is for a single server.

Does lowering variability help as much as adding capacity?

Often it helps a lot. The what-if table in the calculator compares cutting utilization by 10 points with halving service variability for your numbers. Which one wins depends on your numbers, so the table computes both.

Why does the wait explode near 100% utilization?

There is no spare capacity to absorb random bunching. Every extra bit of load adds a smaller amount of idle time to catch up, so the line grows roughly as 1 ÷ (1 − utilization).

Keep going

Updated 17 November 2026. Free to use, no account. The calculation runs in your browser; nothing you type is sent to us.