Real-time numerical simulations of data center air-cooling is achieved using a computational fluid dynamics research code executed on multiple graphics processing units (GPUs). The simulated thermal fields are validated against transient time-series data recorded during the experimental operation of a slab-floor data center that is thermally managed using computer room air handling units.A developed lattice Boltzmann method (LBM) simulation using the Bhatnagar-Gross-Krook (BGK) collision operator is employed to model both the momentum and energy transport equations of fluid dynamics. Airflow turbulence is captured using a large eddy simulation (LES) approach and the effects of natural convection of the air are included using the Boussinesq approximation.The BGK-LBM computation is distributed across 10 GPUs on a multi-GPU remote server. Using optimization strategies for synchronization between GPUs, the computational performance is shown to scale almost linearly with the number of GPUs involved. A parallel algorithm based on MapReduce is developed that can provide continuous measurements of the simulated macroscopic field variables. Agreement between the simulated and measured fields is shown. The numerical simulation can be executed in real-time or faster depending on the lattice resolution.

Achieving energy and cooling efficiency in data center convective air flow and heat transfer can be a challenging task. Among different numerical methods to work with such issues is the Finite Volume Method in Computational Fluid Dynamics. This work evaluates the performance of two such solvers provided by OpenFOAM® in solving this type of convective heat-transfer problem, namely BuoyantBoussinesqPimpleFOAM and BuoyantPimpleFOAM. This is done for two different flow configurations of significantly different Richardson number. To sufficiently resolve the flow, grid sizing effects are elucidated by way of the kernel density estimate. It determines the volume distribution of the temperature in the data center configuration. For the k-epsilon turbulence model used here, it was found that the compressible solver performs faster and requires less grid resolution for both flow configurations. This is attributed to the nature of the boundary conditions which are set such that the mass flow conservation per server rack and cooling unit is achieved. Transient solutions are found to provide better iterative convergence for cases that involves buoyancy, compressibility and flow separation. This is, in comparison to steady-state solutions where artificial numerical pressure drop is found, to depend on the momentum relaxation factors for the convective case with a higher Richardson number.

A multiphysics Simulation-Driven Design approach has been undertaken to augment the OCP Leopard Server thermal management and heat recovery hardware with the Nexalus hybrid liquid-cooled sealed server technology. Independent testing at the RISE Research Institute of Sweden has proven up to 98% heat recovery is achievable at water temperatures up to and exceeding 65°C. The improved design could maintain the elevated water temperature over a range of CPU workloads, ranging from 8% to 75%. Importantly, the design solution achieves this within an architecture that is IOU in height, half that of the original stock 20U server, potentially doubling the compute density of a rack.

Cooling of IT equipment consumes a large proportion of a modern data centre's energy budget and is therefore an important target for optimal control. This study analyses a scaled down system of six servers with cooling fans by implementing a minimal data driven time-series model in TensorFlow/Keras, a modern software package popular for deep learning. The model is inspired by the physical laws of heat exchange, but with all parameters obtained by optimisation. It is encoded as a customised Recurrent Neural Network and exposed to the time-series data via n-step Prediction Error Minimisation (PEM). The thus obtained Digital Twin of the physical system is then used directly to construct a Model Predictive Control (MPC) type regulator that executes in real time. The MPC is then compared in simulation with a self-tuning PID controller that adjust its parameters on-line by gradient descent.