The new reporting scheme adopted by the European Union focuses on the efficiency assessment of data centers based on key performance indicators (KPIs). This has the potential to lead to further measures for improving energy efficiency of data centers. Consequently, a systematic approach is crucial for assessing and improving data center performance. This study investigates efficiency assessment of two European pilot data centers, located in Denmark and Poland, according to the KPIs calculated through Computational Fluid Dynamics (CFD) simulations of airflow and thermal structures. Two separate experimental approaches were adopted at the pilot sites. One measured airflow and temperatures using external sensors at server inlets and outlets, while the other measured inlet temperatures from embedded server sensors. The good agreement observed between simulated and experimental data confirmed the accuracy of the computational model. The validated CFD model was applied to investigate efficiency improvement opportunities by retrofitting the thermal environment such as rack positioning, containment implementation, and guided airflow control. A series of numerical simulations are performed using the validated numerical model and KPIs are calculated for each design. Numerical simulations demonstrate that the efficiency of an existing data center can be enhanced by up to 75% using the presented approach, although the improvement varied significantly with the specific KPIs. The computational approach proposed here can be readily generalized to guide the efficiency assessment and improvement of existing data centers.

Small-scale data centers suffering from low cooling efficiency consume an intense power for reliable operation of IT equipment. In this study, thermal distribution in an air-cooled data center is simulated using an open source Computational Fluid Dynamics (CFD) model to examine the underlying mechanism that reduces thermal and cooling efficiencies. The numerical model is validated with the temperature measurements conducted in the data center. Numerical simulations have revealed that recirculating hot flows observed at the top of the racks increased the air temperature beyond the allowable maximum temperature. To this end, the data center has been retrofitted by creating a hot aisle with the implementation of a moving baffle at the rear of the rack. Numerical simulations conducted for two working scenarios have demonstrated that such a minor modification could result in remarkable enhancement in the cooling efficiency. Efficiency of the data center has improved by 47.2% and 22.7% with respect to the RCI (Rack Cooling Index) and RHI (Return Heat Index), respectively. The present numerical model can capture distributions of the efficiency metrics over the racks. Present methodology can be used to reduce power consumption by the cooling and ventilation systems in existing data centers.

The CFD-based study identified promising retrofit strategies, but numerically optimized designs are not always fully implementable on site. In this follow-up field study, the CFD-recommended actions were physically installed at AAU and validated with real measurements (server inlet/outlet temperatures, airflow velocities, rack-level power) under four IT load levels (25%, 50%, 75%, 100%), so both thermal and energy KPIs could be compared directly against the baseline under real operating conditions.

Averaged across all four tested IT load levels (25-100%).

All three retrofit configurations improved on the baseline, but through different mechanisms. The Plate Test alone reduced RTI sharply (202.02 → 90.49) by closing unused server openings, but RCI_LO actually worsened (178.7 → 212.6), showing local sealing alone doesn't give the most balanced rack inlet temperatures. Plate+Wall gave the best overall result: RCI_HI, RCI_LO, RHI and RI all improved the most, PUE dropped from 3.24 to 2.18 and CoP rose from 0.45 to 0.84 - a 32.7% energy efficiency improvement. Applied to a representative 300 kW small-scale data center, this translates to roughly 2.07 GWh/year in energy savings, about €0.38M/year in cost savings, and about 382 tCO₂e/year in avoided emissions.

PSNC is a micro-data center with two rack cabinets in an open-aisle layout, cooled by ceiling-mounted air conditioners. Unlike AAU, only one retrofitted configuration was field-tested here (not four), because several CFD-recommended actions - notably a thermal separator - could not be installed due to practical and safety constraints. The implemented actions were: rack rotation and repositioning, closing/minimizing ventilation grills, removing one server unit and non-essential furniture, and running the supply air at a controlled 18°C. The results below should be read as a partially implemented retrofit, not the fully CFD-optimized design.

Enhancement (%) is not printed directly in the source report's table for PSNC (unlike AAU's Table 5) - calculated here the same way the source paper computes its own CFD Enhancement figures (% change in the improving direction per KPI).

The largest improvement was in RTI, which dropped from 275.20 to 107.76 - a substantial reduction in cold air bypass and much better balance between supplied cooling air and heat removed from the IT equipment. RCI_LO improved from 64.55 to 86.34, showing more effective cold air delivery to rack inlets, and RHI rose from 32.00 to 47.00, showing better use of the supplied cold air - though still below the ≥80 target, so heat capture wasn't fully optimized. RI improved only moderately (69.52 → 75.26) and RCI_HI barely moved (301.27 → 288.19), staying far from the ideal value of 100. This is consistent with the thermal separator - the measure CFD identified as most impactful - not being implemented here. In short: the feasible field actions clearly helped, but full CFD-recommended airflow separation (as achieved at AAU) would likely be needed to fully resolve recirculation and localized hotspot risk at PSNC.

Interestingly, the original CFD-optimized design (see "Retrofitted → CFD") actually predicted RCI_HI and RCI_LO getting worse (-18.67% and -54.76%) under the fully separated design, likely a side effect of concentrating airflow more tightly once the thermal separator is in place. The field-measured partial retrofit shows the opposite for those two KPIs (+4.34% and +33.76%) - a reminder that the two results represent different designs (full CFD-optimized vs. partially field-implemented) and are not directly comparable one-to-one. Both approaches agree that RTI, RHI, and RI all improve.

The retrofit brought RCI_HI to the ideal 100%, a +22.90% improvement, while RCI_LO was already at the ideal 100% before the retrofit and remained there. RTI improved by 9.10%, moving closer to the ideal 100% mark (from 110.36% to 101.26%), indicating better balance between supplied cooling air and heat removed from the IT equipment. RHI improved by 10.38%, showing better use of the supplied cold air. SHI showed only a marginal change (+0.10%), as it was already close to its target range before the retrofit.