Data centres (DCs) provide critical infrastructure underpinning modern economies; they hold the software and data that modern life depends on. These facilities use massive amounts of power; from a few kilowatts up to hundreds of megawatts, much of it being generated in traditional carbon producing power stations. Although there is an increasing trend towards green energy supply, the need to improve DC efficiency as a whole is critical to UK industry. To date, most of the effort has focussed on physical systems. This project will radically improve current approaches by greatly expanding on the methodologies used model and simulate servers. We will introduce a step-change to DC efficiency by creating a holistic framework that accounts for software, hardware, facility, and human behaviours and use it in training advanced intelligent agents to achieve substantial energy reductions without affecting performance. The number and size of DCs is growing rapidly as they are the backbone of emerging technologies like IoT, 5G, AI, etc. Current DC energy usage is approximately 3% of global consumption; but could reach as high as 10% by 2030. To remain competitive in this growing sector it is vital that UK DCs keep their energy usage in check to compete with regions where green power is cheap. DC efficiency is driven by a number of factors: reducing carbon emissions, reducing costs and increasing capability where power is restricted. Edgetic is an early stage technology company aiming to improve DC efficiency via software services. The standard measure of DC efficiency is PUE (Power Utilisation Effectiveness): a ratio of the power consumed by the whole facility to that consumed by the IT equipment. A PUE of 1 is a theoretical minimum implying energy is only used by the IT hardware; efficiency worsens as PUE increases. Focusing on PUE, the industry has prioritised improving isolated peripheral systems rather than reducing overall energy consumption. PUE improvements are slowing as peripheral, co-dependant systems reach the limits of individual optimisation; improving IT efficiency is the next research frontier. Edgetic uses predictive mathematical modes of IT behaviour to make optimising decisions for the DC. However, our current approach requires individually modelling each workload and type of server in a DC. At present this is acceptable, but in order to substantially grow the business it is vital to improve the scalability of the modelling process since every DC is unique. Every additional variation in hardware and workload substantially increases the required evaluation. The aim of this project is to develop novel methods to speed up server evaluation, estimate behaviours of new hardware combinations and predict performance for different workloads. Uniquely, these methods will be employed in both the existing optimisation technology and provide the foundation for new artificial intelligence tools to optimise DC operation using holistic behaviour simulations. The holistic approach will allow automatic DC optimisation using new operating strategies tailored to individual DCs based on their required characteristics. This has the benefit of radically improving data centre efficiency which in turn reduces the climate impact of DCs and maintains the UK's leading position in the data centre industry.

Resource scheduling in massive-scale distributed systems is the process of matching demand with supply. Demand is associated with requests for resources to execute workloads, such as jobs, tasks and applications. Typical resources in a distributed computing system include servers within a data centre cluster. A scheduler aims to achieve several goals, for example, to maximise system throughput, to minimise response time, to optimise energy usage, etc. These goals may conflict (e.g. throughput versus latency), and the scheduler needs to make a suitable compromise, depending on the user's needs and objectives. In a data centre system with hundreds of thousands of distributed servers, its massive scale is characterised by a number of factors that contribute to the system complexity: - the number of server nodes in the cluster, interconnections between resources and heterogeneity of resources (different types of CPUs, memories, local storages); - the number of concurrent jobs in the system and their arrival rate; - heterogeneity of jobs (different requirements of CPU, memory and local storage; different patterns of resource usage, long-running jobs vs short-alive jobs; urgent jobs vs jobs with loose deadlines). The key requirement for the system is its scalability - the ability of the system to sustain the required throughput level (such as operations per second) while confining the perceptional response latencies to a level similar to a small or medium size system. In our project, we aim to address the following challenges: (a) scheduling at scale (to make prompt scheduling decisions at a rapid rate); (b) resource utilisation at scale (to improve utilisation of resources while maintaining high quality of service); (c) Quality-of-Service provision at scale (to satisfy requirements of diverse workloads). Existing scheduling algorithms developed for practical systems are often designed largely based on empirical knowledge, experience, and best effort. Due to the lack of theoretical foundation, performance of those algorithms cannot be always guaranteed. On the other hand, scheduling algorithms proposed by the theoretical community are usually based on oversimplified abstract system models. Theoretically sound algorithms, with guaranteed accuracy and time complexity, are often impractical because system models do not reflect practical complexity of real systems, and even minor adjustments of system models towards real systems make algorithms no longer applicable. In our project, theoretical and applied experts will consolidate efforts to conduct jointly an interdisciplinary study, overcoming the shortcomings of isolated research. Overall, our project is 1) methodologically driven, attempting to extend the applicability of the most powerful techniques of mathematical optimisation; 2) application driven, where the challenges of massive-scale distributed systems invoke new developments of scheduling methodology; and 3) practice driven, where the research direction is based on hands-on experience of distributed systems specialists.