Internal CFD of server rooms.

Why simulate the inside of a server room?

Internal CFD simulation of server rooms is the historic core of EOLIOS: more than 350 data centers studied across 20 countries. We work on every type of room, from the hyperscale data hall to the enterprise or colocation room, in design as well as on existing sites, which we audit and model. Because an IT room’s cooling can be correctly sized in terms of power and yet fail locally: everything comes down to the distribution of the air.

Global thermal balances guarantee that the installed cooling power covers the dissipated power. They say nothing about what matters most: the temperature of the air that actually enters each server. Between the cooling unit and the rack, the air mixes, bypasses and recirculates; two neighbouring racks can draw in air 10 °C apart. It is this invisible path that CFD simulation makes visible.

What internal CFD simulation secures

The airflow of a server room

The reference layout separates the flows into cold aisles, where fresh air is supplied in front of the racks’ front faces, and hot aisles, which collect the air exhausted at the rear. The raised-floor plenum, the perforated tiles and the CRAC/CRAH units close the loop.

This separation is never perfect. Two phenomena degrade it, and these are exactly what CFD quantifies precisely.

The thermal reading: bypass, recirculation and hot spots

The pressure reading: the cold-aisle / hot-aisle differential

A few pascals are enough to tip a room from one regime to the other, and that is the whole point of the adjustment: too little pressure on the cold side and hot air flows back to the intakes, too much pressure and cold air short-circuits the servers to return straight to the return. The gauge below sums up this reading, the one our engineers apply from the very first calculation, unit by unit and tile by tile.

What the CFD model reconstructs

The model reproduces the room as it is, or as it will be: geometry, equipment and operating regimes. It then solves the fluid-mechanics equations over millions of cells to restore the velocity, temperature and pressure fields at every point.

The quality of the result depends on the fidelity of these inputs. On an existing room, a measurement campaign (temperatures, flow rates, thermography, smoke tests) makes it possible to calibrate the model: the digital twin first reproduces the observed state, then serves as a reliable test bench for changes. This step is by no means mandatory: on a new project, the model works directly from the drawings and the manufacturer data.

“The airflow of a raised-floor grille can vary by more than 50 % depending on its position, the room configuration and its distance from the air-conditioning systems.”

ASHRAE, CAT, SIT: the criteria that count

A room is not judged “on average”. The CFD standards of the major colocation and hyperscale operators rest on two normalised temperatures, which every simulated case must respect.

The exact thresholds vary from one operator to another and depend on the cooling mode: stricter for chiller-based architectures, widened for free cooling and evaporative cooling. All of them fit within the public ASHRAE TC 9.9 framework: the recommended range of 18 to 27 °C, then the allowable ranges of classes A1 to A4 that permit controlled excursions. In every case, compliance is judged with N cooling units, redundancy being consumed, on humidity as well as temperature; in transient, a higher tolerance applies, but strictly timed.

The restitution by SIT bins

The report does not stop at a minimum and a maximum: each rack is classified by SIT bin, case by case. This reading shows at a glance how many racks are compliant, how many live on the margin and where the exceedances are concentrated.

Scenarios: the cascade of cases, from nominal to the worst failure

The nominal regime

The first calculation establishes the reference mapping: inlet air temperature of each rack, flow rates actually delivered by each tile, pressure balance. It reveals the existing hot spots and prioritises the corrections: tile adjustment, blanking of free U slots, flow rebalancing, containment, down to fine tuning such as grille placement, CRAC/CRAH setpoints or the chilled-water temperature.

The cascade of cases from the hyperscale standards

The CFD specifications of hyperscale and colocation operators leave no choice of scenarios: they impose a cascade approach, where each case inherits from the previous one and answers a precise question. Our studies follow it natively.

Failure scenarios

A room compliant in nominal regime can tip over within minutes as soon as a piece of equipment stops. The simulation first assesses common operational failures: loss of one or more cooling systems, shutdown of a unit for maintenance, closure of a damper. The search for the worst case is systematic: a unit is not removed at random, several combinations of shutdowns are re-simulated, including electrical failures, since a single switchboard can stop several units at once. On the calibrated model, we check that the remaining air conditioners hold the load without any rack leaving the allowable range, identify the most exposed rows and concretely validate the stated redundancy (N+1, 2N) rather than assuming it.

Upgrade projects

Densification, arrival of high-power AI or HPC racks, switch to containment, raising of setpoints to improve the PUE: each variant is tested virtually on the calibrated model. Trade-offs are made on compared temperature maps, not on rule-of-thumb figures. An upgrade campaign does not stop at the IT room, moreover: the additional load feeds through to the UPS, the LV switchboards and the batteries, whose cooling comes under the thermal study of the technical premises. The same model finally allows the works to be sequenced: which racks to move first, which units to stop during the migration, and in which order, without ever leaving the allowable range.

The internal study finally ties in with the external CFD simulation of the data center: recycling of hot-air plumes on the roof and power losses of the dry coolers in extreme conditions directly condition the cooling capacity available in the room. The standards even impose it explicitly: the internal cases are simulated under the worst scenario of the external CFD, incorporating the resulting derating of the cooling-production equipment.

Power cut: the minutes that count

The most critical scenario is handled in transient regime: what happens during a power cut on the site? The calculation follows the temperature rise minute by minute and quantifies the time available before reaching the servers’ critical thresholds. It accounts for the inertia of the chilled-water loop, the switchover sequence to UPS then generators and the restart of the backup equipment, pinpointing the most unfavourable servers. This objective delay becomes the key input for the operating procedures and the sizing of the backup systems.

Nothing here is a flat rate. Each source of thermal inertia is justified: chilled-water loop volume, storage buffers, mass of the slab and the racks. The restart times of the chillers and generators come from the factory acceptance tests (FAT) when they have taken place; failing that, the simulation is re-run as soon as the tests deliver the real timings. And when the site has a continuous backed-up cooling, the transient study then serves to size its inertia rather than to demonstrate its necessity. These cooling reserves are themselves the subject of full CFD studies: the stratification of the thermal-storage tanks conditions the energy actually available at the moment of the cut, and the hydraulic behaviour of the chilled-water loop and its networks sets the speed at which this cooling reaches the rooms.

An often-forgotten criterion completes the analysis: the temperature reached in the hot aisle is compared with the trigger threshold of the sprinkler heads, to check that an electrical cut cannot degenerate into an untimely release of the fire extinguishing above the racks.

Containment, high density and liquid cooling

Physical aisle containment is the most structuring lever for separating the flows. The choice of configuration depends on the room, its height, its air-return mode and its operating constraints: CFD makes it possible to compare the options on your real geometry.

In every case, the real efficiency depends on the leaks: missing panels, cable passages, empty racks. The simulation takes them into account and avoids overestimating the gain of a theoretically perfect containment.

Containment is also a financial lever: it can qualify for energy-saving certificates under the scheme dedicated to aisle containment in computer rooms. A documented CFD study, with quantified before/after gains, is a solid piece of evidence to support the application.

Accommodating high density and DLC in an existing room

Integrating high-density or liquid-cooled racks (DLC, Direct Liquid Cooling) disrupts existing rooms and calls for a two-step approach. First a capacity assessment: the simulation projects the resources actually available (electrical power, cooling production, flow rates); it does not invent them. Then the physical constraints the calculation does not address directly: floor load capacity, circulation zones, maintenance operations and machine turnover. These preliminary checks condition the value of the study.

Then comes the “mix cooling” trade-off: making air cooling, DLC and cold doors coexist in the same room. CFD compares the possible distributions and reveals the unsuspected effects: vortices and local depressions, heat dissipation of the busbar trunking, pressure losses of the hydraulic networks.

Hybrid rooms: air and liquid coexist

Liquid cooling does not remove the airflow, it reconfigures it. Part of the heat is still rejected to the air: peripheral electronics of the DLC racks, power supplies, CDU heat exchangers, and the classic racks the standards require to be interleaved between the liquid positions. The specifications therefore require a minimum airflow to be maintained, a fraction of the nominal flow defined by the standard: CFD checks that this residual flow, delivered by fewer units, still reaches every rack that remains air-cooled.

The model incorporates the CDUs, the hydraulic networks running in the hot aisle above the racks with their per-branch load limits, the air-cooling units deactivated to free up space, and the resulting clutter in the plenum. The air/liquid distribution of each row is arbitrated on temperature maps, position by position, including for the future deployments the room will have to absorb.

The EOLIOS method

EOLIOS is a consultancy of engineers specialised in the climatic design of data centers. For us, internal CFD is not a visualisation exercise: it is a design tool, grounded in a daily practice of IT rooms.

Our teams design and make server rooms reliable for hyperscale, colocation and enterprise infrastructure players, in France and internationally. We work across the whole life cycle: climatic sizing from the sketch stage, aisle layout, redundancy validation in an extreme scenario, commissioning with load banks, then operational support through the digital twin. This dual culture, design and operation, lets us translate every simulation into realistic action plans, compatible with operational practices and service continuity. Investors, integrators and operators rely on our studies to arbitrate their climatic architecture choices and document their availability commitments.

Because the tool is not everything: without an operator’s reading, a simulation remains an image. Every study leads to concrete, directly applicable recommendations (control strategies, degraded modes, monitoring indicators), prioritised according to two constant priorities: service continuity and energy efficiency.

Our studies natively meet the CFD specifications of hyperscale and colocation operators: imposed cascade of cases, CAT/SIT criteria, minimum report contents, deliverables in English when the standard requires it. And because a project cannot wait: our dedicated computing power lets us run urgent simulations, with first results in a few days when the client need demands it.

The results take the form of 3D images and animations understandable by all the project stakeholders, from investor to operator, usable both in meetings and in technical reports. When confidentiality requires it, the models can only be read through a dedicated viewer supplied by EOLIOS.