In JMeter's Gaussian Random Timer, is the Deviation field a cap on the extra pause?

## What the field actually multiplies The Gaussian Random Timer stores the same two properties as its uniform sibling — `ConstantTimer.delay` for **Constant Delay Offset** and `RandomTimer.range` for **Deviation** — and differs only in the draw: ```java // GaussianRandomTimer.delay() return (long) Math.abs((getRandom().nextGaussian() * getRange()) + super.delay()); ``` `nextGaussian()` produces a bell-shaped value centred on 0 with standard deviation 1. Multiplying by the Deviation field stretches that bell; adding the offset shifts it. So Deviation 200 with Constant Delay Offset 1000 gives pauses centred near 1000 ms, with roughly two draws in three inside 800–1200 ms — and a small share outside that window in either direction. It is a *spread* control, not a limit. ## Compared with the other two random timers | Timer | Random field | Shape of the draw | Is the field a bound? | |---|---|---|---| | Uniform Random Timer | Random Delay Maximum | flat over 0 to the field | yes, it is a hard ceiling on the random part | | Gaussian Random Timer | Deviation | bell centred on the offset | no, it is a standard deviation | | Poisson Random Timer | Lambda | Poisson-shaped, mean Lambda | no, it is the mean of the draw | The Poisson timer is the closest cousin in intent: `PoissonRandomTimer.delay()` is `Math.abs(randomPoisson((int) Math.round(getRange())) + super.delay())`, so **Lambda** is rounded to a whole number of milliseconds and used as the mean of the Poisson component, which is then added to the offset. Note that Lambda is rounded — a Lambda of 0.4 rounds to 0 and the timer degenerates to the Constant Delay Offset. ## The absolute value is not cosmetic For the uniform timer, `Math.abs` can never change anything: a non-negative offset plus a non-negative draw is non-negative. For the Gaussian timer it can. Set Deviation 500 and Constant Delay Offset 100 and a draw of −1.5 gives −650 ms, which JMeter turns into a 650 ms pause. The consequence is that with a small offset relative to the Deviation you no longer get a bell centred on the offset — you get a folded distribution whose mass piles up near zero and whose mean is higher than the offset. If you want the shape to behave as the field names suggest, keep the Constant Delay Offset comfortably larger than the Deviation. ## Practical checks - **Read the units.** Both fields are milliseconds, and both are saved as plain strings: the GUI writes the field text straight into `RandomTimer.range`, so a typed `100` is stored as `100`. The `100.0` you often see in a `.jmx` is the GUI's own default value for that field, left untouched. - **There is no maximum field.** If you need a hard ceiling on the pause, the Uniform Random Timer is the element that has one; the Gaussian timer cannot express it. - **Non-numeric values collapse to zero.** Both fields go through JMeter's string-to-number conversion, which returns `0` when parsing fails instead of raising an error, so a typo or an unresolved reference degrades the timer silently rather than failing the plan. - **`timer.factor` scales it.** The Gaussian, Uniform and Poisson timers all implement `ModifiableTimer`, so if the `timer.factor` property is set to anything other than `1.0f` the computed delay is multiplied by it before the thread sleeps. - **The draw is per thread.** `RandomTimer.getRandom()` returns `ThreadLocalRandom.current()`, so each thread has its own generator and threads do not serialise on a shared one. ## How this is usually probed Interviewers rarely ask the formula outright. They ask why a run configured with Deviation 100 and offset 100 produced pauses of 400 ms, or why the observed average pause is higher than the Constant Delay Offset. Both answers come from the same two facts: the Gaussian tail is unbounded, and the absolute value folds the negative side of the bell back onto the positive side.