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.

This paper analyzes the prospects of a holistic air-cooling strategy that enables synchronisation of data center facility fans and server fans to minimize data center energy use. Each server is equipped with a custom circuit board which controls the fans using a proportional, integral and derivative (PID) controller running on the servers operating system to maintain constant operating temperatures, irrespective of environmental conditions or workload. Experiments are carried out in a server wind tunnel which is controlled to mimic data center environmental conditions. The wind tunnel fan, humidifier and heater are controlled via separate PID controllers to maintain a prescribed pressure drop across the server with air entering at a defined temperature and humidity. The experiments demonstrate server operating temperatures which optimally trade off power losses versus server fan power, while examining the effect on the temperature difference, ∆T. Furthermore the results are theoretically applied to a direct fresh air cooled data center to obtain holistic sweet spots for the servers, revealing that the minimum energy use is already attained by factory control. Power consumption and Power Usage Effectiveness (PUE) are also compared, confirming that decreasing the PUE can increase the overall data center power consumption. Lastly the effect of decreased server inlet temperatures is examined showing that lower inlet temperatures can reduce both energy consumption and PUE.

Finding synergies between heat producing and heat consuming actors in an economy provides opportunity for more efficient energy utilization and reduction of overall power consumption. We propose to use low-grade heat recovered from data centers directly in food processing industries, for example for the drying of fruit and berries. This study analyses how the heat output of industrial IT-load on servers can dry apples in a small-scale experimental set up.To keep the temperatures of the server exhaust airflow near a desired set-point we use a model predictive controller (MPC) re-purposed to the drying experiment set-up from a previous work that used machine learning models for cluster thermal management. Thus, conditions with for example 37 C for 8 hours drying can be obtained with results very similar to conventional drying of apples.The proposed solution increases the value output of the electricity used in a data center by capturing and using the excess heat that would otherwise be exhausted. The results from our experiments show that drying foods with excess heat from data center is possible with potential of strengthening the food processing industry and contribute to food self-sufficiency in northern Sweden.

Low latency requirements are expected to increase with 5G telecommunications driving data and compute to EDGE data centers located in cities near to end users.

This article presents a testbed for such data centers that has been built at RISE ICE Datacenter in northern Sweden in order to perform full stack experiments on load balancing, cooling, micro-grid interactions and the use of renewable energy sources. This system is described with details on both hardware components and software implementations used for data collection and control. A use case for off-grid operation is presented to demonstrate how the test lab can be used for experiments on edge data center design, control and autonomous operation.