# Kingman sensitivity: what the calculation means

This is a worked evaluation of the VUT approximation, not a simulation or a
storage benchmark. The six cases hold mean service at 1 ms and service SCV at
one, while varying utilization and interarrival SCV. Values in the JSON retain
exact rational arithmetic; the article uses three significant digits.

## Reproduce

Use Python 3.10 or newer; no third-party libraries are needed.

```sh
python3 -B model.py --check
python3 -B -m unittest discover -s . -p test_model.py
python3 -B -O -m unittest discover -s . -p test_model.py
```

Run `python3 -B model.py` to regenerate results.json. Its source hash binds the
calculation. Inputs are exact decimal/rational strings, integers or Fractions;
floats and booleans are rejected. The date records this publication's actual
generation date, not a claim about external data freshness.

## Read the rows

The evaluated expression is ((arrival SCV + service SCV)/2) * rho/(1-rho) *
mean service. SCV means variance divided by the square of the mean. It is
dimensionless. Arrival rate is 800 or 950 requests/s; mean service is 1 ms.

One admissible family has iid gamma interarrival times with shape 1/arrival_SCV
and scale E[A]*arrival_SCV, and independent exponential service. This specifies
a stationary GI/G/1 FCFS model with unlimited waiting space, finite moments,
no vacations or abandonment, and rho below one. No random samples are generated.
When arrival SCV is one in this specified family, arrivals are Poisson; the
mean then coincides with the exact M/M/1 result. This special case does not
make the other rows exact queueing results.

The supplementary finite distribution uses intervals 1/2 and 3 with probabilities
4/5 and 1/5. Its mean and variance are both one: SCV one alone does not identify
an exponential distribution. Multiplying every interval by a positive constant
changes its mean and variance but leaves SCV unchanged.

## Limits and evidence

The tests check arithmetic, units, stability, invalid inputs and the algebraic
Poisson special case. They do not validate the approximation's error at any
load. No confidence interval, p99 or hardware performance is inferred.
SCVs summarize marginal variation, not serial dependence. A closed-loop source,
correlated arrivals or parallel SSD resources require a separately justified model.

Formula: Harchol-Balter and Scully, section 1:
https://www.cs.cmu.edu/~harchol/Papers/TeaPACS21.pdf
Heavy-traffic and moment-convergence assumptions: Boon, Janssen and van Leeuwaarden,
arXiv:2206.09844v1, sections 1-3:
https://arxiv.org/abs/2206.09844v1

All numeric rows are Hesela-authored constructed examples. No third-party paper,
figure or empirical dataset is redistributed.
