Three settings you may never have opened can change which racks pass: the turbulence model, the order of the numerical scheme, and the point where you stop iterating. On the halls we have compared, switching these moved the result about as much as, or more than, refining the mesh. That is not a flaw in CFD. It is a reason to know what each setting does and to check your own result against it.
Turbulence: how the model treats mixing
Air in a data hall is turbulent, and turbulence mixes hot air into cold air. The solver cannot resolve every eddy, so it uses a model for the mixing.
A mixing-length model is the simplest. It sets the eddy mixing from the distance to the nearest wall, with no extra equations to solve. It is fast and steady, and it works well when flow is mostly aligned and attached. It does not know about turbulence that is carried along by the flow, which matters behind racks, at hot-aisle exits and in recirculating regions.
A k-ε model solves two extra equations, one for the energy in the turbulence (k) and one for how fast it dissipates (ε). Turbulence is then produced where the flow shears and carried downstream. It costs more per iteration and needs more care to run, and it generally handles jets and recirculation better.
Order: how sharp the scheme is
A first-order scheme is stable and smooth, and it blurs. Where a cold jet meets hot air it spreads the boundary over several cells, which tends to mix more than the real air does. A second-order scheme keeps the boundary sharper, and with it the peaks that decide pass or fail.
Blurring has a direction. A first-order result tends to under-report the hottest spots, because it mixes them away. On an example cold-aisle containment room we solve for our own studies, the hottest rack inlet read 79.9 °F (26.6 °C) at first order, which is 0.7 °F under the 80.6 °F (27 °C) recommended limit. A second-order continuation of the same run put it at 81.0 °F (27.2 °C), 0.4 °F over. The room did not change. The answer to “does every rack pass?” did.
On customer halls we have studied, the same switch has put several more racks over the allowable limit. Changing the turbulence model from mixing length to k-ε has moved the count the other way, taking racks back under the limit. The two changes pushed in different directions, which is the point: neither setting is a free choice.
When to stop: a stopped run is not a result
A solver iterates until a rule says to stop. The rule might be a residual target, an iteration cap, or a time budget. Each can stop a run that has not settled, and then the numbers you read are a snapshot of a room still moving.
We call a run settled when two things are true: the residuals are flat, and the readings you care about have stopped changing. For a data hall those readings are the rack intake temperatures. Residuals alone can mislead, because the equations can look quiet while a recirculation cell is still slowly warming one end of a row.
On the same example room, the rack-intake probes had settled by iteration 540 at a quick mesh and 790 at a standard one, ahead of stops at 570 and 1,120. Running further changed the hottest rack inlet by 0.1 °F or less, and no rack average by more than 0.2 °F. That is what a good stop looks like: you can show that going on would not have changed the report. What kept moving was the air speed at the probes, which took 950 and 1,515 iterations to flatten, a reminder that you should decide which quantity you are waiting for.
The failure case is the one where the answer drifts. A run that keeps swinging by a degree or more at an intake when you cut it off has not produced a temperature. It has produced a place to restart.
How to check your own result
Three checks cost little and catch most of the trouble.
- Rerun at second order, starting from the first-order solution. If the hottest inlets move by a degree or more, report the second-order numbers or both.
- Plot every rack intake temperature against iteration. If any line still has a slope or a swing at the end, the run has not settled, whatever the residuals say.
- Solve three meshes, coarse, medium and fine, and compare the hottest rack inlet as well as the averages. Averages agree early. Hot spots do not, and they are what you are judging.
If the three agree, you have a result. If they do not, say which setting each difference came from before anyone reads a pass or fail off it.
What we do
When the physics permits, our engineers solve at second order with k-ε, check intake temperatures over the iterations rather than the residuals alone, and report a mesh ladder where the decision is close. Some models need a different treatment. A nearly closed tile or containment panel can make a run fail outright when it is modeled as an ordinary porous layer, so our solver models tiles and panels 10 percent open or less as thin resistance plates instead.
To see how we check results before they reach a report, read how we know it’s right.