At present, data centres (DCs) are one of the major energy consumers and source of CO2 emissions globally. The GREENDC (Green Data Centres) project addresses this growing challenge by developing and exploiting a novel approach to forecasting energy demands.
The project will bring together five leading academic and industrial partners with the overall aim of reducing energy consumption and CO2 emissions in specific national DCs. It will implement a total of 163 person-months staff and knowledge exchanges between industry and academic partners. More specifically, knowledge of data centres operations will be transferred from industry to academic partners, whereas simulation based optimization for best practice of energy demand control will be transmitted from academia to industry through the knowledge transfer scheme.
In recent years, the use of the information and communications technologies (ICT), comprising communication devices and/or applications, DCs, internet infrastructure, mobile devices, computer and network hardware and software and so on, has increased rapidly. Internet service providers such as Amazon, Google, Yahoo, etc., representing the largest stakeholders in the IT sector, constructed a large number of geographically distributed internet data centers (IDCs) to satisfy the growing demand and providing reliable low latency services to customers. These IDCs have a large number of servers, large-scale storage units, networking equipment, infrastructure, etc., to distribute power and provide cooling. The number of IDCs owned by the leading IT companies is drastically increasing every year.
As a result, the number of servers needed has reached an astonishing number. According to recent reports, the approximate number of servers large companies possess are as follows: Google (≈1 million), Microsoft (>1 million), Akamai (≈127K), INTEL (≈100K), facebook (>300K), ebay(≈60K), Rackspace (≈100K). With rapid increase in demand for internet service providers, the energy used by IDCs has been skyrocketing due to the large number of hosted servers and associated workload. The global energy cost for enterprise and DCs power and cooling reached $30 billion in year 2008 only. According the EPA report published in 2007, IDCs consumed about 61 billion kilowatt-hour (kWh) in 2006 for a total electricity cost of about $4.5 billion.
This means that IDCs consumed 1.5% of total U.S. electricity consumption. It was projected that this would rise to 3% in 2011 or in other words about 120 billion kWh equivalent to the average consumption of a city with 11.6 million households. In reports published in 2013 it is shown how the IT sector is consuming about 10% of the world’s electricity generation. According to JRC, IDCs consume already about 50 TWh, which is projected to increase to 100 TWh by 2020. DCs electricity consumption in Western Europe alone was 56 TWh in 2007 and is projected to rise to 104 TWh in 2020. Due to large amount of energy implied and the related cost, IDCs can make a significant contribution to the energy efficiency by reducing energy consumption and power management of IT ecosystems.
This is why most researchers focus on reducing power consumption of IDCs. One of the ways to reduce power consumption of IDCs is to optimise load-balancing among servers to prevent unnecessary overloads to part of the servers that leads to higher power consumption. Traditionally, electric energy systems are unidirectional in operation and top-down oriented. Demand and supply balance is kept by feeding the system with large power plants (sized to hold the peak request but only for short times). This balance is crucial and should be kept for all times, and also, is under the threat of intermittent nature of renewable energy sources (whose presence and quantity largely fluctuates in time) and increasing number of the electric vehicles (which act as either capacitors or load for the grid depending on their status).
Thus new challenges faced by power systems operators requires new tools and sophisticated control methodologies. The idea of using the load as a control medium has been adopted by the “Utilities” for a long time as a mechanism for power curve smoothing.
As a result, we have seen the birth of the concept of demand-side management (DSM), which refers to the desired changes in the timing and amount of the electricity demand by applying load control systems, improved energy-efficiency cycles, on-site energy storage packages and the promotion of off-peak usage of electricity. Fast acting DSM tools have only recently become affordable due to the development of accessible and cheap global communication infrastructure.
In addition, embedded systems now make it relatively easy to add a certain dose of “smartness” to the loads themselves. The development in this area is driven by the fact that—despite increased efficiency of electric devices—consumption is steadily rising some percent every year due to the increase in numbers of customers (i.e. both at industrial as well as household or commercial level). While generation might not be much of a problem (it is always possible to build new production facilities), it is the grid capacity that makes many of the involved stakeholders concerned. Most of the losses occur during energy conversion and transport, which are functionalities of the Grid, and additionally there is the problem of limited resources, demand growth, as well as needs for production configurations that can take into account the long term sustainability factor. However, existing studies in optimised load-balancing have limitations in the basic assumption on the relationships between losses and workloads.
Existing studies and commercial solutions mainly aims at optimal distribution of work-loads between the servers. They generally assume a linear relationship between losses and workload. Thus, they propose that the best solution is simply loading of the servers up to a certain limit and then move the remaining load to the next available one. Troublingly, however, the losses are not linear due to jet effect. High load may cause higher losses. Rather than using the simple linear model, this project will use a quadratic equation model instead. Also, the current literature only considers load optimisation among the servers. That is an optimisation in the space dimension.
The GREENDC project will optimise the systems time dimension as well. The project will take a more holistic view by considering the system as a whole i.e. servers, cooling system, backup power and electrical distribution system. Here is an example of using this approach: For a unit of workload, an IDC create heat and need to cool it down. Let us assume one unit of workload requires one unit of cooling power. If we do both at the same time we need two units of power and this creates losses square of that power i.e. 4 times. If we instead overcool the system before the workload and allow slight overheat and again overcool the system after the workload, we can distribute the load in time. That means doing the same work in the longer term. That halves the losses. In our approach we will distribute the workload in space and time by utilising thermal inertia of the system as an energy storage medium.
Finally, the GREENDC project will develop an IDC simulator in which IDC managers can evaluate different strategies to minimise energy consumption and CO2 emissions in the consideration of space and time dimensions as described above.
The secondments proposed in this project will allow for the above factors to be studied through simulation techniques and algorithms, resulting in important knowledge transfers between the IDCs, consortium partners, industry and academia, and through outreach activities create more awareness about the importance of these challenges for the public to take action into pushing the involved stakeholders to reduce the ecological footprint.
The specific objectives of the GREENDC project are as follows.
- defining energy demand forecasting and control problems through a systematic review of the literature on the energy optimisation in IDCs considering carbon footprint and field studies in two real world IDCs at industrial partners to develop a deeper understanding of the key variables and parameters that affect GHG emissions, service level and cost of IDC operations;
- mathematical and simulation modelling of IDC operations and developing an efficient and scalable metaheuristic optimisation techniques to guide the search in an interactive mode with the simulation module for trade-off analysis in the form of Pareto-optimal frontier;
- prototyping a decision support tool (GREENDC DSS) for industrial applications by industrial partners (TSAT and DAVID) and other companies operating IDCs;
- testing, validating and benchmarking of the decision support tool using case studies from TSAT and DAVID that ensure minimum 10% improvement in energy consumption of the pilot IDCs in Turkey and Bulgaria;
- dissemination of results to practitioners as well as academic beneficiaries through conferences, workshops and academic publications meeting minimum threshold
- providing a new and lasting collaboration opportunities between UBRUN, TSAT, GTU, DAVID and LKKE through 215 PMs of cross-sectorial knowledge exchange by 16 experienced and 13 early-stage researchers; and
- training early stage researchers through their participation in the research and technological development activities of the project.