Modern data centres require control, which aims to improve their energy efficiency and maintain their high availability. This work considers the implementation of a server fan agent, which is intended to minimise the power consumption of the corresponding server fan or group of fans. In the paper, the reinforcement learning approach to energy-efficient control of server fans is suggested. The reinforcement learning workflow is considered. The Simulink blocks simplifying the building of the environment for the reinforcement learning agent are developed. This work provides the framework for creating and training reinforcement learning agents of different types. As the paper is only a work-in-progress, possible type of agents and their training process is described, but training and deploying the agent is a work for the future.

Modern data centres consume electricity at the industrial scale; at the same time, most of them demonstrate redundancy in energy consumption. The two most significant energy consumers in a data centre are its computational system and cooling system. This work focuses on techniques, which adapt the system energy use based on resource requirements at run-time. Actually, this work is an inception phase, which determines the main requirements to control the energy-efficient data centre and develops its general project. For that aim, the general design of the multi-agent control is proposed. The different types of agents are identified and their objectives are determined. Based on agent types the architecture of the multi-agent control is developed and agents' interactions are considered. The paper presents an example of an agent controlling a server fan. The agent is examined using closed-loop co-simulation with a server model.

Abstract- The main goal of climate control systems in data centres is to keep the temperature and humidity in a suitable level for computational devices. Therefore, cooling and humidification systems are essential parts of every Building Automation System (BAS), which is utilized in server rooms. Although the current climate control systems ensure appropriate thermal conditions to computational nodes such as servers, they waste substantial amount of energy. The main cause of this inefficiency is that the current climate control systems, which are responsible for thermal management of the data centres, follow rigid control strategies that maintain constant thermal conditions irrespective of climate changes caused by various computational loads in the plant. To address this issue, in our previous works we proposed a method of optimizing energy consumption in data centre cooling systems while maintaining an acceptable level of thermal comfort for CPUs in the server room. In this paper, we present the enhancement of our previous method by incorporating humidity control into it. The enhanced method consists of a thermal model of server room, and a simulation tool to find an energy efficient control strategy for the climate control system in different situations by comparing different control strategies. The effectiveness of the proposed method has been investigated via simulation and the result, which shows 41.5% reduction in total energy consumption is presented.

Contemporary data centres consume electricity onan industrial scale and require control to improve energyefficiency and maintain high availability. The article proposes anidea and structure of the framework supporting development andvalidation of the multi-agent control for the energy-efficient datacentre. The framework comprises two subsystems: the modellingtoolbox and the controlling toolbox. This work focuses on suchessential components of the controlling toolbox, as an individualcontroller. The reinforcement learning approach is applied to thecontrollers’ implementation. The server fan controller, named SFagent, is implemented based on the framework infrastructureand reinforcement learning approach. The agent’s capability ofenergy-saving is demonstrated.

Modern data centres consume a significant amount of electricity. Therefore, they require techniques for improving energy efficiency and reducing energy waste. The promising energy-saving methods are those, which adapt the system energy use based on resource requirements at run-time. These techniques require testing their performance, reliability and effect on power consumption in data centres. Generally, real data centres cannot be used as a test site because of such experiments may violate safety and security protocols. Therefore, examining the performance of different energy-saving strategies requires a model, which can replace the real data centre. The model is expected to accurately estimate the energy consumption of data centre components depending on their utilisation. This work presents a toolbox for data centre modelling. The toolbox is a set of building blocks representing individual components of a typical data centre. The paper concentrates on parameter estimation methods, which use data, collected from a real data centre and adjust parameters of building blocks so that the model represents the data centre most accurately. The paper also demonstrates the results of parameters estimation on an example of EDGE module of SICS ICE data centre located in Luleå, Sweden. 

Data centers as all complex systems are prone to faults, and cost of them can be very high. This paper is focused on detecting the faults in the cooling systems, in particular on local fans level. In the paper, a hybrid approach is proposed. In the approach a model is used as substitute of the real system to generate dataset containing records of both normal and fault cases. On the generated data, machine learning algorithm or ensemble of algorithms are selected and trained to detect the faults. To demonstrate the approach, the rack model of real data center is created, and reliability of the model is shown. Using the model, the dataset with normal as well as abnormal records of data is generated. To detect faults of local fans, simple classifiers are built for all pairs: a local fan – a processor unit. Classifiers are trained on one part of generated data (training data), and then their accuracy is estimated on another part of generated data (test data). A real-time fault detection system is built based on the classifiers. The rack model is used as the substitute of the real plant to check operability of the system.

For most modern data centres, it is of high value to select practical methods for improving energy efficiency and reducing energy waste. IT-equipment and cooling systems are the two most significant energy consumers in data centres, thus the energy efficiency of any data centre mainly relies on the energy efficiency of its computational and cooling systems. Existing techniques of optimising the energy usage of both these systems have to be compared. However, such experiments cannot be conducted in real plants as they may harm the electronic equipment. This paper proposes a modelling toolbox which enables building models of data centres of any scale and configuration with relative ease. The toolbox is implemented as a set of building blocks which model individual components of a typical data centre, such as processors, local fans, servers, units of cooling systems, it provides methods of adjusting the internal parameters of the building blocks, as well as contains constructors utilising the building blocks for building models of data centre systems of different levels from server to the server room. The data centre model is meant to accurate estimating the energy consumption as well as the evolution of the temperature of all computational nodes and the air temperature inside the data centre. The constructed model capable of substitute for the real data centre at examining the performance of different energy-saving strategies in dynamic mode: the model provides information about data centre operating states at each time point (as model outputs) and takes values of adjustable parameters as the control signals from system implementing energy-saving algorithm (as model inputs). For Module 1 of the SICS ICE data centre located in Luleå, Sweden, the model was constructed from the building blocks. After adjusting the internal parameters of the building blocks, the model demonstrated the behaviour quite close to real data from the SICS ICE data centre. Therefore the model is applicable to use as a substitute for the real data centre. Some examples of using the model for testing energy-saving strategies are presented at the end of the paper.

Cooling system is the second most energy-consuming part of a modern data center; herewith it often consumes energy inefficiently. Therefore, any possibility allowing to reduce energy consumption of cooling systems should be studied and put into practice if successful. In this paper, the idea about impact of IT load distribution between servers on energy consumption of data center cooling system is proposed. The idea is in distribution of total IT load between servers according to the location of the servers inside racks. To test idea, the model simulating the thermal behaviour and energy consumption of a real data center located in Northern Sweden is developed. Comparing the data obtained from the real plant and the data generated at the simulation shows the reliability of the model. Two strategies of IT load distribution between servers are considered, and their impact on energy consumption of cooling system is demonstrated.