Textbook presentations of queueing theory start by assuming that the time between customer arrivals and the time to serve a customer are both exponentially distributed.
Maybe these assumptions are realistic enough, but maybe not. For example, maybe there’s a cutoff on service time. If someone takes too long, you tell them there’s no more you can do and send them on their merry way. Or maybe the tails on times between arrivals are longer than those of an exponential distribution.
The default model makes other simplifying assumptions than just the arrival and service distributions. It also makes other simplifying assumptions. For example, it assumes no balking: customers are infinitely patient, willing to wait as long as necessary without giving up. Obviously this isn’t realistic in general, and yet it might be adequate when wait times are short enough, depending on the application. We will remove the exponential probability distribution assumption but keep the other assumptions.
Assuming exponential gaps between arrivals and exponential service times has a big advantage: it leads to closed-form expressions for wait times and other statistics. This model is known as a M/M/s model where the two Ms stand for “Markov” and the s stands for the number of servers. If you assume a general probability distribution on arrival gaps and service times, this is known as a G/G/s model, G for “general.”
The M/M/s model isn’t just a nice tractable model. It’s also the basis of an approximation for more general models.
Let A be a random variable modeling the time between arrivals and let S be a random variable modeling service times.
The coefficient of variation of a random variable X is its standard deviation divided by its mean. Let cA and cS be the coefficients of variation for A and S respectively. Then the expected wait time for the G/G/s model is approximately proportional to the expected wait time for the corresponding M/M/s model, and the proportionality constant is given in terms of the coefficients of variation [1].
Note that the coefficient of variation for an exponential random variable is 1 and so the approximation is exact for exponential distributions.
How good is the approximation? You could do a simulation to check.
But if you have to do a simulation, what good is the approximation? Well, for one thing, it could help you test and debug your simulation. You might first calculate outputs for the M/M/s model, then the G/G/s approximation, and see whether your simulation results are close to these numbers. If they are, that’s a good sign.
You might use the approximation above for pencil-and-paper calculations to get a ballpark idea of what your value of s needs to be before you refine it based on simulation results. (If you refine it at all. Maybe the departure of your queueing system from M/M/s is less important than other modeling considerations.)
[1] See Probability, Statistics, and Queueing Theory by Arnold O. Allen.