M/G/1 queue calculator: average wait for one server with any service-time spread

An M/G/1 queue has random arrivals, one server and a general service-time distribution. The Pollaczek-Khinchine formula gives the average wait from only three numbers: arrival rate, mean service time and the coefficient of variation of service. Enter them below. Constant service halves the wait of exponential service at the same load.

Constant service, 80% busy2.00 minM/D/1, cv 0: the wait is half the M/M/1 wait
Exponential service, 80% busy4.00 minM/M/1, cv 1
Long-tailed service, cv 210.00 min2.5 times the M/M/1 wait; simulated 10.44 min

How to calculate the wait in an M/G/1 queue

  1. Enter arrivals per hour and the average service time in minutes to get utilization.
  2. Measure the standard deviation of service times and divide it by the mean to get cv.
  3. Enter cv: 0 for constant service, 1 for exponential, above 1 for a long tail.
  4. Read the average wait, line length and time in system.
  5. Press the simulation check to compare the formula with a seeded run.

Your single server

Starting values are an example, not data about your process. Measure the mean and the standard deviation of your service times, then divide the second by the first to get cv. Utilization must stay below 100%.

Result

Utilization80%
Average wait in line2.5 min Pollaczek-Khinchine
Average line length2 Lq = λ × Wq
Time in system3.5 min wait + service
Average wait for the same arrivals and mean service time at different service-time variability
cv of serviceCaseAverage wait
0M/D/1, constant service2 min
0.5More regular than exponential2.5 min
1M/M/1, exponential service4 min
1.5More spread than exponential6.5 min
2More spread than exponential10 min
Open this exact case in Simflowly

Exact for random arrivals, one server, one unlimited line and any service-time shape with the cv you give. The simulation uses a lognormal shape for cv other than 0 and 1.

The formula

ρ = λ × E[S] (utilization, must be below 1) Wq = ρ × (1 + cs²) ÷ (2 × (1 − ρ)) × E[S] Lq = λ × Wq (Little's law) W = Wq + E[S] (time in system)

E[S] is the mean service time and cs its coefficient of variation. The only thing that depends on the shape of service is the factor (1 + cs²) ÷ 2: it is 0.5 for constant service, 1 for exponential and 2.5 at cs = 2. The wait scales with it directly.

Worked example: one server, 80% busy

48 arrivals an hour at random, a mean service time of 1 minute, so ρ = 0.8. Only the spread of service changes. Formula against the engine (seed 7, 30 replications of 200 hours, lognormal shape except cv 0 and 1):

Average wait for one server at 80% utilization by service-time variability, formula and simulation
cv of serviceCaseFormulaSimulated (95% interval)
0Constant (M/D/1)2 min2 min (1.9 min to 2.1 min)
0.5More regular than exponential2.5 min2.5 min (2.4 min to 2.6 min)
1Exponential (M/M/1)4 min3.9 min (3.8 min to 4.1 min)
1.5Long tail6.5 min6.7 min (6.2 min to 7.2 min)
2Long tail10 min10.4 min (9.1 min to 11.8 min)

The formula sits inside the simulated interval on every row. The constant and exponential rows are the same two results used in the engine validation: 2.0000 against 2.0044 simulated for M/D/1, and 4.0000 for M/M/1.

Why variability costs waiting time

Two servers can be equally busy and have very different lines. With constant service every customer takes exactly the mean, so the line is only fed by random arrivals. With a long tail, a few very long services block the server while others pile up behind them. At the same 80% load, cv 2 gives 10 min where constant service gives 2 min. Reducing the spread of service, by standard work, by splitting long jobs or by routing them elsewhere, cuts the wait without adding capacity.

When this model is wrong

  • Arrivals are not random. Scheduled arrivals are more regular and waits are lower; waves and batches make them higher. The formula for variable arrivals as well is on the wait vs utilization calculator.
  • The system is not in steady state. A queue that starts empty needs a long run to reach the average, and a system over 100% busy has no average at all.
  • Customers leave or jump. Abandonment, priorities and several servers are different queues.

Questions

What is the Pollaczek-Khinchine formula?

It gives the exact average wait in an M/G/1 queue: Wq = ρ × (1 + cs²) ÷ (2 × (1 − ρ)) × mean service time, where ρ is utilization and cs is the coefficient of variation of service. It needs random arrivals, one server and an unlimited line, and nothing else about the service distribution.

What is the difference between M/M/1, M/D/1 and M/G/1?

All three have random arrivals and one server. M/M/1 assumes exponential service times (cv 1), M/D/1 assumes constant ones (cv 0), and M/G/1 lets you choose any shape through its cv. M/M/1 and M/D/1 are special cases of M/G/1 with cv 1 and cv 0.

Where do I get the coefficient of variation?

Time a sample of real service times, compute their mean and standard deviation, and divide the second by the first. Do not assume it: measure your own. Use a sample large enough that the standard deviation is stable before you trust the wait.

Does it work with several servers?

No. This formula is exact only for one server. For several servers with exponential service use the Erlang C queue calculator, and for several servers with other service shapes use a simulation or an approximation. The M/G/c case has no simple exact formula.

Keep going

Updated 20 November 2026. Free to use, no account. The calculation runs in your browser; nothing you type is sent to us. The engine behind the simulation checks is validated against exact queueing results.