Stability Testing
How Long Should a Burn-In Run?
There is no documented number of hours that certifies a machine. There are reasons to run longer, and a point where the extra hours stop paying.
No software page this desk consults publishes a duration for a burn-in. The honest answer to how long is a length chosen for a reason you can state, and this page lays out the reasons, because the figure itself does not exist.
The figure no documentation prints
A forum answers this question with bare numbers, stated like specifications. The software's own download page stays structural: such a page states what the tool is for, what it runs on, and which version it is. Hours are not a property of the tool. They belong to the question you are asking with it, which is why no tool documents them.
That absence is the first finding. A machine rebuilt after a part swap and a machine trusted to run unattended through a long weekend are not asking the same question, so they cannot share a number. A duration quoted without the machine, the workload, and the room attached to it is practice repeated as a specification. Practice is worth something; this desk only asks that it be labeled.
What do the first minutes establish?
The opening stretch of a run proves that the load runs. The tool starts, the machine accepts the work, and the faults that were never marginal announce themselves at once, because a fault that large does not need hours to surface. A failure that arrives in the first minutes is not bad news; it is fast news, and it usually points somewhere specific.
What those minutes cannot give you is the machine as it will live under load. A run begins on hardware still carrying idle temperatures, and a pass, like every pass, is a sentence about conditions. The first minutes interview a machine that is about to stop existing.
What do the extra hours buy?
Three mechanisms, each a reason rather than a number. Heat: a machine presents its worst self only after warmth has soaked through every part, and a run shorter than that soak tests a cooler machine than the one you will live with. Recurrence: an error that appears rarely cannot be excluded by any run shorter than its own rhythm, and that is arithmetic, not opinion. Variety: a long workload is rarely one load but a sequence of internal stages, and a longer run meets more of them.
| Stretch of the run | What it can honestly say | What it says nothing about |
|---|---|---|
| The first minutes | The load starts and runs; the machine is not grossly broken. | Rarer faults, and anything about a machine warmed through. |
| The soak, once readings stop climbing | Behavior at the machine's settled heat, on that one workload. | Other workloads, other rooms, other seasons. |
| Long unattended stretches | The machine sustains the load through hours nobody watches. | Whether it always will; no pass faces forward. |
| Any completed run | Those patterns, that speed, that heat, that day, produced no detected error. | Everything outside that sentence. |
The workload decides what the hours cover
Hours measure exposure, not breadth. The subsystem a load saturates defines what the run samples, so equal durations on different loads answer different questions, and a long run on one load narrows the claim instead of widening it. Reading the subsystem matrix before choosing is the difference between asking a question and hoping.
Some loads make the stopping decision explicit, because they have no finish line of their own. GIMPS, the Mersenne prime search, offers its software free for Windows, Linux, FreeBSD and Mac OSX, and its download page describes the point of the exercise as contributing the processing power a machine is not otherwise using. A workload built on that premise has no finish line the page publishes, which is also why it gets borrowed for bench work: when the work never ends, the run stops because you decided, and the decision needs a reason of its own.
When is a run long enough?
Long enough for what. A rebuild that only has to prove it survived assembly can get its answer early, from the minutes that catch gross faults. A machine that has to hold a sustained load through hours nobody watches needs a run containing stretches that resemble those hours, logged with its temperatures so that unattended never means unwatched, and no shorter argument substitutes for them.
No duration certifies stability; what a stress run proves is bounded by its own conditions, and the boundary is the honest part. So the gesture that closes this subject is written before the run starts. Write down the sentence you want to be able to say afterwards: this machine, that load, that stretch, that room, no detected error. Run until the sentence is true; the number was never the part that held up.
Before the clock starts
- Name the question the run exists to answer; a run without a question cannot be long enough.
- Record the workload, the start time, and the conditions of the room, because the pass sentence needs its clauses.
- Let the machine warm through under load before trusting temperature and voltage sensors that look calm on a cold start.
- Decide in advance what counts as failure, a reset, a reported error, a freeze, so the verdict is not improvised at the moment of failure.
- Plan to treat a failure as information about where to look, not as a verdict on the part you already suspected.
Where this goes wrong
The common drift is adding hours to a single workload and calling the machine stable. More hours on one load extend the claim in time and leave it unchanged in kind: the pass still describes that load, for longer.
The quieter error is carrying an old conclusion across a change. A driver update, a moved module, a different season in the room, each one restarts the clock on what the earlier run said, and reusing a past pass as a standing property of the machine spends one careful diagnosis twice.
mersenne.org
The download page of GIMPS, the Mersenne prime search; the page title also carries the name PrimeNet. It presents the project's software, Prime95, as free, for Windows, Linux, FreeBSD and Mac OSX, and describes running it as a way to contribute a machine's spare processing power. The version listed when this desk read the page was 30.19 build 20.