# Same mean service, different mean waiting

These are constructed analytic M/G/1/FCFS examples, not storage measurements.
One server, independent Poisson arrivals, iid service requirements, unlimited
waiting room, no abandonment or vacations, stationary regime and utilization
below one. Durations are milliseconds; arrival rates are requests/second.

Both distributions have mean service 1 ms. The mixture assigns 99/100 of
requests 1/2 ms and 1/100 of requests 101/2 ms. Its second moment is 103/4 ms^2,
not the square of its 1 ms mean. At 200 requests/s (20% utilization), mean
queueing is 103/32 ms instead of 1/8 ms for constant service: 25.75 times as
large. At 800/s the respective means are 103/2 ms and 2 ms. The long class is
1% of requests but 101/200 = 50.5% of service work. No empirical uncertainty
interval is appropriate for these exact chosen inputs.

Run Python 3.10+ standard library in this folder:
`python3 -m unittest test_model.py`
`python3 model.py --check`

JSON retains exact rational strings and display numbers. Units, assumptions
and the SHA-256 of model.py travel with the results. Floats, booleans, invalid
probabilities, negative rates and unstable queues are rejected.
The zero-rate row in tests is a limiting case, not an observed arrival sample.

Formula: Larson and Odoni, Urban Operations Research, section 4.7, Eq. 4.81:
https://web.mit.edu/urban_or_book/www/book/chapter4/4.7.html
Interpretation: Harchol-Balter and Scully, section 1:
https://www.cs.cmu.edu/~harchol/Papers/TeaPACS21.pdf

This mean formula does not yield p99. Real device parallelism, correlated
arrivals, feedback, vacations and scheduling need a matching model.
