Site selection for a data centre – a multi-criteria decision-making model

Abstract

IT data centres (DCs) are a huge investment and most mid-size companies need to set up their own DCs to run their business operations. Building new DCs is a time-consuming effort needing millions of dollars and months of planning. Companies cannot afford to make a wrong decision when designing and setting up a new DC as these need to build for a life of around 20 years. In this paper, we studied the challenges and issues faced by enterprise DCs and explored various criteria which can be used for selecting a site for a new DC. We use a hierarchical model to look into various financial, environmental, social, political and geographical factors which an organisation need to look into when choosing a city for their DC. The model discussed in the paper provides a general framework and can be adopted by a company operating in any industry and any country.

Keywords:

• energy conservation and low-carbon manufacturing
• engineering design for sustainable development
• sustainable technology innovation

1. Introduction

Most of the businesses run their operations based on information technology processes and data centres (DCs) form the heart of IT infrastructure of any organisation. For all medium and multinational organisations, DCs form the backbone of critical business processes. Variety of services such as web hosting, database applications, telecom services and banking services depend on the DC infrastructure to offer services to clients. With the advent of e-commerce, smart phones, social networking and mobile communications, requirements of enterprise-wide DCs keep on increasing.

Companies want to build DCs in locations which are not prone to natural disasters, terrorism which and also provide cheap and reliable resources such as electricity, network and transport facilities. At the same time, they need a robust facility which can be operated 24 × 7 without disruptions and provide scalable computing power.

2. Literature review

A data center (or data centre or datacenter or datacenter) is a facility used to house computer systems and associated components, such as telecommunications and storage systems. It generally includes redundant or backup power supplies, redundant data communications connections, environmental controls (e.g., air conditioning, fire suppression) and security devices. (Wikipedia 2011)

A DC set-up needs investments in different areas – building construction, massive power and cooling requirements, cabling for power, network, storage, telecom, etc., computing equipment such as servers, switches, routers, racks and storage frames. DCs are a massive investment for many businesses. Only very large companies can afford to set up their own DCs. Some medium-size companies have to rely on the hosting or cloud services offered by IT vendors. A new DC investment can run into millions of dollars and take 2–3 years to plan and build. This paper looks at a decision model which can be adopted by companies worldwide to choose a location for their enterprise DCs.

Many offices and industries started using computers in the 1980s. The main purpose of computers was to provide computing power and automation not achieved by manual methods of working. The 1990s saw an increase in the complexity of information and a demand for a more controlled environment for the IT systems. Client–server computing became the buzz word and this type of computing required different types of servers such as web servers, app servers, email servers and database servers which required to be installed in specialised rooms with controlled humidity and temperature. In the early 2000s, we saw fast growth of Internet which led to exploding growth in the number of DCs. Next century will be the century of cloud computing and social networking. Services such as Google Apps Facebook, YouTube and Netflix will drive the further growth of DCs. As content will be stored centrally instead of physical media such as DVDs and cassettes, we will need more servers and storage frames. Most IT professionals understand the connection between energy efficiency and long-term DC sustainability, but many have trouble bridging the gap between the ideas of efficient design and actually implementing it in their infrastructure. As DCs are the backbone of business processes for enterprises, meeting the required service-level agreement for infrastructure is very critical for the IT organisations. DCs need to operate 24 × 7 × 365 and getting a downtime for maintenance is extremely difficult. Therefore, IT managers need to plan the facilities and equipment design keeping these requirements in mind.

A recent survey (NREI 2009) of senior decision makers at large corporations in North America found that 83% of respondents are planning DC expansion in the next 12–24 months, and 73% of respondents plan to add two or more facilities as part of their DC expansion. Moreover, DC and IT budgets are both projected to increase by 8% in 2010, up from 7% and 6%, respectively, last year. ‘At the end of the day, a DC is the lifeblood of corporate America. If a DC goes down, a company is dead in the water’.

As per Ellison and Sanchez (2009), a company's DC strategy should not be limited to consolidation of DCs. The goal of the strategy should be to optimise DC infrastructure to respond faster to business needs while enhancing the services and value IT brings to the business. Companies can measure the performance of their DCs on some key metrics such as efficiency, utilisation, performance and capacity.

In response to high growth of DCs in developed nations, Canadian Standards Association (CSA) and American National Standards Institute (ANSI) have jointly developed a standard for DC infrastructure. It is called ANSI/CSA/EIA/TIA 942. It is developed in collaboration with Electronics Industries Alliance (EIA) and Telecommunications Industry Association (TIA). The purpose of this effort was to provide standards for planning of DCs, computer rooms, co-location centres, trading floor equipment rooms, technology test laboratories and similar spaces (Diminico 2011). Majority of the technical content deals with facility specifications. It provides guidelines on a wide range of subjects, useful to someone designing or managing a DC. It will be an official tiering standard for determining the quality of a DC and for comparing DCs with each other.

The Green Grid (TGG) is a global consortium of companies, government agencies and educational institutions dedicated to advancing energy efficiency in DCs and business computing ecosystems. With input from representatives around the world, Singh et al. (2011) have developed the DC Maturity Model (DCMM), which provides clear goals and direction for improving energy efficiency and sustainability across all aspects of the DC. The DCMM touches upon the major components of the DC including power, cooling, computing, storage and network. The levels of the model outline current best practices and a 5-year roadmap for the industry. The DCMM provides capability descriptions by DC area such that operators can benchmark their current performance using the DCMM equaliser, thereby determining their levels of maturity and identifying the ongoing steps and innovations required as part of their DC and IT strategy to achieve greater energy efficiency and sustainability improvements, both today and into the future.

Currently, there are lots of debate in the IT industry about the evolution of DCs – mainly with the implementation of cloud services and exponential growth of online data. One of the foremost among them that continues to continuously be debated about and continuously evolving would be the DC metrics (Daim et al. 2009). Companies need to understand the importance of energy efficiency metrics in order to examine the effectiveness of their green initiatives for energy savings in DCs.

Different DC metrics which are developed by TGG and Uptime Institute (Pfleuger 2008) are as follows:

•	Power usage efficiency.
•	Water usage effectiveness.
•	Energy reuse effectiveness.
•	DC compute efficiency.
•	DC performance per Watt.

All these metrics help the DC stakeholders in the usage of different resources so that they can better plan for these limited resources.

Corporate average DC efficiency (CADE) is another metric used to rate the overall energy efficiency of an organisation's DCs. CADE was introduced in a joint repot from the Uptime Institute and McKinsey (Kaplan et al. 2008) that proposed the metric as a single key performance indicator that could be used to compare the energy consumption of one DC against that of another.

3. Problem statement

Growth of DCs will not happen without problems. Companies are facing a growing set of challenges in the DCs. Traditional measures of operational capability such as system availability and resource utilisation are facing increasingly stringent expectations, and new metrics such as energy efficiency have been added to the mix. At the same time, because of IT's increasing critical role in business, DCs are facing a new set of external demands such as global availability, regulatory compliance and the expectations of the empowered user (Kaplan et al. 2008).

DCs are consuming more financial resources and emitting more green gases. As per McKinsey & Company Report – there are three distinct issues in DCs.

Server utilisation in a typical DC is at 6% – facility utilisation can be as low as 50%. Application developers do not fine-tune their applications to use minimum resources. Business managers want real-time data without knowing how it impacts DC costs, and infrastructure managers choose cheap equipment without caring for energy efficiency of the equipment. Idle servers typically consume one-half of the electricity of fully utilised servers, but there is a push today to develop truly energy-proportional equipment, in which the electricity consumption more closely matches the server utilisation (Karidis and Watson 2010).

DCs face new pressures to reduce costs, energy consumption and physical footprint. As a result, many have turned to server consolidation and server virtualisation to address their pain points.

Server virtualisation has proven to reduce DC capital and operational expenses over and over in the industry, but this assumes a consolidation ratio that will offset the cost of the virtualisation software licencing as per Jones (2009). A simple calculation that assumes a server cost of $8.5 K, virtualisation software licence cost of $3 K both amortised over 3 years, plus power, maintenance and administrative costs, will yield a capital plus operating cost reduction of nearly threefold if the server workload consolidation ratio is 10:1.

DCs are the biggest consumers of energy after power plants. There are two main areas of energy usage in a DCs:

a.	cooling (including humidity control) and
b.	hardware equipment used for computing.

Cooling expense can be the biggest operating expense in a DC. There are different types of cooling – such as hot aisle/cold aisle and evaporative cooling. Cooling expenses are such a big concern that many companies are planning their DCs in cold countries such as Finland and Siberia to use weather-based cooling. As per Data Center Report (2011), HP, IBM and Dell have all set up their DCs in Finland.

Energy consumption and thermal requirements are fast becoming the limiting factors in DC environments. The cooling infrastructure is a significant part of a DC. The complex connection of chillers, compressors and air handlers creates the optimal computing environment, ensuring the longevity of the servers installed within and the vitality of the organisation they support. Increasing server densities and workloads pose power and cooling demands that could require DC operators to renovate existing facilities or even build new facilities (Sharma et al. 2004).

As per EMC's report on data centers (Syed 2008), the key objective when designing an optimal DC cooling system is to create a clear path of air flow from the source of cooled air to the intake of equipment, and from the hot air exhaust of the equipment to exhaust system's return air duct.

IT has experienced a real rush to ‘go green’ and the reasoning is easy to understand – by lowering the energy use of computing equipment, a company can reduce its impact on the environment and save money on power costs in the process.

One of the easiest and most straightforward approaches to power conservation is to select and deploy Energy Star-rated systems, which have expanded to embrace enterprise-class servers.

Energy Star is a voluntary labelling programme developed in 1992 by the US Environmental Protection Agency (EPA) and the US Department of Energy to ‘protect the environment through energy efficient products and practices’.

Buying the right IT equipment for the right price. Earlier, DC managers only used to think about the cost of the equipment. Now with energy prices increasing rapidly, they need to strike a balance between the equipment cost and the energy requirement of the IT equipment. Selection of hardware (servers, routers, switches and frames) is a major factor dictating the cost and energy consumption. Energy Star compliant hardware is expensive to buy, but the energy savings over the life of the equipment outweigh the cost difference.

Given the astonishing growth in DC power demands, the EPA has developed and released an initial version of the Energy Star Computer Server Specification (Olsen 2011).

A single kilowatt does not sound like much in a DC until you factor in that it is consumed continuously (the proverbial 7/24/365), which adds up to 8760 kWh per year. At 11.5 cents per kWh, 1 kW costs $1000 per year. (Of course, 11.5 cents is just an average, and in many areas the cost is much higher). Over a 3-year period, that ‘small’ 500 W server can cost $3000 or more just in energy consumption. In fact, since many of these small servers cost < $3000, we can see why some analysts have predicted that the power to operate a server will exceed the server's price, especially as the cost of energy rises (Syed 2008).

DCs are now the biggest consumers of IT budgets – 25% of IT budgets is spent on DCs. Moreover, accountability for DC costs is fragmented. It is spread between facilities and IT budgets – this makes the finding total cost of ownership (TCO) even more difficult. End users of DC services do not have any visibility into the costs.

All these inefficiencies in the design and operation of DCs can impact company's bottom line directly.

Growth of Cloud is driving lot of interest in DC investments. Companies are using applications sold by cloud providers. They are also working on consolidation of DCs. Even the federal government has come out with a plan to close 137 DCs out of a total of 2000. This move will save the government ∼18 billion dollars per year (Federal Govt. Report 2011).

Utilisation, energy efficiency and high capital investments are main areas of concern for companies in their DCs. Growth of Cloud is driving an interesting debate as to how the future DCs will look like. Companies are also trying modular DCs such as DC in a container.

In order to solve these issues of inefficiency with DCs, companies have multiple options when they need to host their DCs.

1.	Outsource all IT infrastructure.
2.	Co-locate your IT equipment in a leased DC.
3.	Move your applications to Cloud Services.
4.	Set up your own DC.

According to Pund-IT Research an IT industry analysis firm, the following are the issues to be considered behind the cost issues that need to be looked at while determining the need or the continuance of a DC: (King 2011) facilities, equipment, staffing and DC alternatives.

Most major enterprises such as banks, insurance companies, federal agencies, high-tech and healthcare companies need to set up their own DCs due to legal, ownership and business continuity reasons. How do these companies decide on the location of a DC? Even companies such as Accenture, Amazon and Teramark which provide hosting services for other clients need some sort of criteria to select a site for their IT DCs.

In our literature review, we could not find any academic literature which discusses a good decision-making model for DC site selection. As DCs become more and more critical for a company's business, they need to make logical decisions about their DC strategy.

Our paper tries to develop a model for DC site selection using proven methods of decision-making.

Before we delve into the details of different criteria and alternatives, we provide a brief summary of multi-criteria decision-making (MCDM) using analytic hierarchy process (AHP) and hierarchical decision model (HDM).

4. Decision models

The MCDM model, by coupling theory and knowledge, provides an analytical approach to expert consultation and is adapted for a variety of technology and business fields aiming at suitability assessments (Gurumurthy and Kodali 2008).

Mathematical models have been developed extensively with the availability of relatively inexpensive personal computing horsepower. Operations research methods of optimisation can maximise performance, minimise cost or target any criterion of interest. Location research often uses optimisation methods to choose between alternatives. Although mathematical models often employ various quantitative inclusions such as algorithms, yet the overall decision is more based on judgement and politics than on numerical methods (Ajgaonkar et al. 2007).

MCDM methods can handle both quantitative and qualitative criteria and analyse conflict in criteria and decision makers (Pohekar and Ramachandran 2004). Several classification and categorisation exist but in general MCDM methods can be divided into two categories: multi-objective decision-making (MODM) and multi-attribute decision-making (MADM: Climaco 1997). In MODM, the DM problem is characterised by the existence of multiple and competitive objectives that should be optimised against a set of feasible and available constraints (Diakaki et al. 2009) rather than as in MADM evaluating a set of alternatives against a set of criteria. MADM is one of the most popular MCDM methods to be adopted to solve problems associated with different problems (Wang et al. 2009). They contain several different methods, the most important are AHP, PROMETHEE, ELECTRE and multi-attribute utility theory (MAUT: Cristóbal 2011). The comparison of MCDM methods related to energy planning is discussed in the literature (Hobbs and Horn 1997, Pohekar and Ramachandran 2004, Polatidis et al. 2006, Zhou et al. 2006, Chu et al. 2007, Wang et al. 2009). In a previous analysis by Pohekar and Ramachandran (2004), MAUT was the most common MCDM method used in energy planning literature, AHP, PROMETHEE, ELECTRE, MAUT, fuzzy methods and decision support systems.

Among MADM models, hierarchical expert judgement quantification models are used for the elucidation of the level of preferences of decision alternatives through judgements made over a number of criteria (Dezert et al. 2002). The major benefit of these models is that a decision maker can utilise it to capture the vague preferences in mind and to sort them into a prioritised sequence systematically. When the problem involves a large number of criteria and alternatives, typical models become very slow and many errors can occur. Therefore, some studies have developed what is called a self-regulating approach of the pairwise comparison (Niknafs et al. 2010).

To make a decision in an organised way to generate priorities, we use a process called AHP which was developed by Saaty (1990).

AHP implementations have been successfully done in a variety of applications in the past. AHP methodology has been used by Kayikci (2010) in Freight industry, by Wang et al. (2011) in evaluation of enterprise development, by Tummala et al. (1997) in Electronics Industry of Hong Kong, by Cheng (1997) in evaluation of missile systems, by Lijuan and Shinan (2011) for evaluation of Human factors in aircraft accidents, by Lin (2010) in evaluation of web sites, by Rajput et al. (2011) in selection of wafer fabrication process, by Durán (2011) in selection of Computer Management Systems, by Tam and Tummala (2001) in vendor selection of Telecom systems, by Bertolini et al. (2006) in proposal selection of public contracts, by Kuo et al. (1999) in location selection of convenience stores, by Ramanathan and Ganesh (1995a) for Resource Allocation problems, by Al Khalil (2002) in selection of project delivery methods, by Gerdsri and Kocaoglu (2007) to build a framework for Technology Roadmapping, by Lai et al. (2002) for software selection, by Korpela et al. (2007) in Warehouse Operator selection, by Rad et al. (2011) in rankings of University Programs, by Ramanathan and Ganesh (1995b) in evaluation of energy alternatives for household lightings, by Wang et al. (2009) in Banks loan decision-making process, by Büyüközkan et al. (2012) in selection of personal digital assistant device, by Kim et al. (2010) for prioritising of emerging technologies and by Bottero et al. (2011) in the assessment of Water Treatment Systems.

HDMs are used in several different decision problems: in the past, these models have been successfully used by Bohanec and Zupan (2004) in real estate, by Bohanec et al. (2000) in healthcare, by Cakir and Canbolat (2008) in inventory management, by Al-Subhi Al-Harbi (2001) in project management, by Azadeh et al. (2011) in railway system improvement, by Bozbura et al. (2007) in human resource management, by Durán and Aguilo (2008) in evaluation and justification of an advanced manufacturing system, by Galan et al. (2007) in manufacturing and by Cunningham and Lei (2007) in selection of new technologies.

5. Model development

In this paper, we use the HDM methodology to develop the decision model.

In this process, we need to decompose the decision into some steps. These steps are outlined as: (1) define the problem and determine the kind of knowledge sought. (2) Structure the decision hierarchy from the top with the goal of the decision, then the objectives from a broad perspective, through the intermediate levels to the lowest level – which usually is a set of the alternatives. (3) Construct a set of pairwise comparison matrices. Each element in an upper level is used to compare the elements in the level immediately below with respect to it. (4) Use the priorities obtained from the comparisons to weigh the priorities in the level immediately below. Do this for every element. Then for each element in the level below add its weighed values and obtain its overall or global priority. Continue this process of weighing and adding until the final priorities of the alternatives in the bottom-most level are obtained (Figure 1).

Figure 1 Model development.

In a hierarchical model, structure of the hierarchy not only depends on the type of the problem, but also on the expertise of the participants involved in the process by Chen et al. (2011). One of the most important steps of this method is to build a matrix A, where each element A ij (i, j = 1, 2,…, n) represents the relative importance of the criterion i over the criterion j. To express this relative importance, the decision maker can make use of a verbal scale. The latter is then transformed into a fundamental scale of absolute numbers taking integer values between 1 and 9 (Verly et al. 2010).

We met with many experts from different departments of the company to gather their inputs about the model development. Thinking from engineer's technical perspective, we were concentrating a lot about the technologies of a DC. Once we started speaking to the experts from finance and operations, we realised that technology factors cannot be mapped to a location. It is not possible to give weights to technological criteria in relation to a city. So, we had to remove the technology factors from the model.

We also researched many papers about the energy efficiency of the DC. Our research showed that energy efficiency in DCs is a complete separate topic of research itself and cannot be tied directly to site selection. Many methods such as brainstorming, interview and nominal group discussion were used to establish a set of criteria and factors through the uses of expert opinions.

TGG (http://www.thegreengrid.org) – which is a consortium of professionals from DC-related companies – also talks about similar factors to be considered when designing sustainable DCs. White Papers written by experts from different IT companies such as Cisco, EMC and Intel were also studied to help in the design of the model. Based on authors' own knowledge, interviews with panel of experts and literature review, the following main criteria were chosen for the HDM.

Based on all the expert inputs and literature review, following main criteria were developed for the HDM.

1.	Geographical
2.	Financial
3.	Political
4.	Social

Each one of the criterion above has several factors that are associated with it. The factors associated with each criterion directly influence the importance and priority of that criterion. Figure 2 shows the model criteria, all factors linking to their criterion and the suggested alternatives for the purpose of applying the model in the following section.

Figure 2 DC site selection HDM.

6. Criteria discussion

Geographical Factors (C1) include the majority of the factors in the model. These are discussed below.

6.1 Disaster avoidance (F11)

Reducing the physical risk of catastrophic facility failure starts with site selection. According to Uptime Institute (2011), DC downtime caused by geographically predictable natural threats (earthquakes, hurricanes, lightning storms, Tornadoes, floods, etc.) are not ‘Acts of God’ but are, instead, management errors.

According to the Federal Emergency Management Agency, 70% of all power outages are caused by weather-related events (Uptime Institute Report on Data Centers 2011). Beyond the direct physical threat posed by natural disasters, DCs can be crippled by damage to supporting infrastructure or rendered inoperable because key employees and vendors are unable to reach the facility. When deciding where to locate a DC, chief information officer (CIOs) must choose an area with a low risk of natural disasters.

Companies can take many steps to build DCs to withstand natural disasters. To prevent flooding in DCs, build them above sea level with no basements, tightly sealed conduits, moisture barriers on exterior walls and dedicated pump rooms; drainage/evacuation systems; Rack systems can be secured to the DC floor to prevent tipping over during tremors and earthquakes using plinth kits.

6.2 Transport/accessibility (F12)

Although DCs can be set up in small cities, these should be still easily accessible by any means of transport. Suppliers of IT equipment and other construction crew need to be able to travel to the site easily. According to reports by Search Data Center Site (2011) – ‘Cities that have a precedent for data centers, vendors have already gravitated to that area’.

6.3 Network carrier availability (F13)

Communications are at the core of a DC functioning. All equipment in a DC needs to communicate with devices spread around the world for different business processes. Companies need fibre connectivity. Location should have more than one network carrier for redundancy. AT&T, Teramark and Verizon are some of the major carriers in the nation and planners need to make sure that enough bandwidth is available for current and future requirements.

Google set up its DC in Dalles (Oregon) due to availability of cheap power and easy access to dark fibre.

6.4 Availability of power (F14)

DCs consume power equal to 2500 homes. Power requirements have been rising from the DCs and this is the reason that EPA has been asked to look into energy efficiency of DCs. With their enormous requirements of power hungry equipments, DCs emit as much carbon dioxide as the nation of Argentina.

Moreover, a typical DC costs in US ∼$200 million. DC electricity costs have become the second highest expense in the DC operations at 13%. In 2005, businesses paid about 20% more for electricity than they did in 2004. As of 2006, DCs used 61 billion KWh of electricity – twice the energy consumed in 2000 equal to output of about 15 thermal power plants (Syed 2008).

Therefore, DC energy efficiency is one of the hot topics in IT industry today.

When considering a green DC initiative, we need to look into inventory of all IT assets to assess and understand current power usage patterns. For increased assessment control and accuracy, consider logical division of the DC into more than one section, where power readings can be made in each section. These logical divisions can be made along the lines of service to business units; an assessment can be performed for a complete infrastructure serving a given business unit. Similarly, a DC asset division can be made by asset type such as servers, networks or storage resources. The goal of an assessment is to identify inefficiencies in existing power and areas of opportunity for the greatest impact.

6.5 Availability of water (F15)

Although not very obvious, DCs also consume huge quantities of water, mainly for the cooling chillers (Schutter 2011). Therefore, setting up DCs in hot climate locations is not desirable. Water is the main reason that some of the new DCs of Google are next to a river in Oregon. Companies also need to look into cost of the water as locations near the rivers will generally have cheap water.

Financial factors (C2) are equally important as DCs are massive investments. Companies cannot afford to make wrong decisions as millions of dollars are at stake.

6.6 Land cost (F21)

Since DCs are massive facilities, cost to acquire the land is a crucial factor in the site selection. Usually land is cheaper in small remote cities. Land in Manhattan is 10 times the cost of same land in a remote area of Oregon/Washington.

6.7 DC building construction cost (F22)

A DC building is one of the major investment and companies cannot afford to make wrong choices on the type and design of buildings as it can impact the overall operations. This is a one-time decision as it is not easy to upgrade the building of a 24 × 7 DC. Therefore, this is a critical factor in the analysis.

According to Brickhouse (2010), ‘A sustainable building is one that uses material and processes aimed at minimizing the facility's impact on both human and environmental health’. Majority of the energy efficient ideas can be implemented when designing and building a DC. Once a building has been constructed, it is very difficult to make changes to power, cooling and other infrastructure. Therefore, majority of energy efficiency decisions needs to be taken at the time of building a DC. IT managers who wish to design sustainable DCs have access to a wide range of strategies. Leadership in energy and environmental design (LEED) certification process offers a useful framework for organising those strategies. LEED is a standard for environmentally sustainable construction created by the US Green building Council, a non-profit organisation that promotes energy efficient architecture.

6.8 Operational cost (F23)

Operational cost consists of the following main components: (1) electricity cost, (2) water cost, (3) power and cooling equipment maintenance cost, (4) IT hardware maintenance cost and (5) manpower cost.

Political factors (C3) impacts all investment decisions even if the operation of a DC is a purely technical issue. All investment decisions need to be reviewed by local Governments. – So political criteria discussion is also equally important.

6.9 Tax structure, incentives and subsidies (F31)

Corporate tax structure varies in different states. As DCs involve massive capital investments, companies look for ways to reduce their taxes. States offering lower taxes appeal to companies interested in making huge capital investments in DCs. For example, Washington State has one of the lowest corporate taxes in the nation. No doubt that Amazon, Microsoft, Yahoo, Providence have set up their DCs in different cities of Washington (Data Center Location Rankings 2011). Also, cities offer tax rebates on capital investments in capital equipment. Personal property taxes on the millions of dollars worth of equipment are generally the next largest operating cost. Having a lesser impact on total operating cost is real property tax on the building. Both of these vary from municipality to municipality.

Offering subsidies and incentives to attract businesses is very common. Mainly small cities offer all sort of incentives and tax breaks for big businesses as those generate jobs and bring in revenue for the state and city.

6.10 Talent pool/jobs creation (F32)

When state and city governments give permits for industries, they look at how the facility will help the community in the long run. ‘Every politician runs on the platform of jobs’. One of the main criteria companies look for is the job creation in the long run. DCs create many jobs during construction. They also need site engineers – electrical/telecom engineers to operate the site. Therefore, IT managers need to consider the local talent pool when looking at a potential DC site.

Social factors (C4) although not related to industries are becoming important as citizens become more environment conscious.

6.11 Safety and security (F41)

Crime rate of the city needs to be checked. Also, terrorism threat needs to be evaluated for the city being considered for building the new DC.

Physical safety of the data centers is of extreme importance as any disturbance to the facility can impact business processes. DC access should be controlled and regularly monitored. World-class DCs have different types of security mechanisms – this includes security officers, Close Captioned TV (CCTV), security cameras, mantraps, audits of visitors and entrance separate from office areas (World Class Data Center Features Report 2011).

Reliable fire protection systems should be installed. DCs host critical business data and data integrity cannot be compromised. Non-water-based fire protection systems are preferred.

6.12 Rules and regulations for urban planning (F42)

What are the environmental and business laws of the various cities? DC is not a very environmentally friendly business due to its huge energy consumption. What is the public view on big businesses? Are there enough industrial districts in the city? These are some of the questions DC planners need to get answers before deciding on a city for setting up DCs. Zoning laws vary from city to city. In some cities, DCs are even allowed in basement of high-rise buildings, while in the other small cities, DCs are located in remote areas away from the residential communities. Permits to build DCs are required from local municipalities – which can be time-consuming. Some cities such as Portland are highly environment conscious and people oppose the idea of any building/industry which harms the environment. Therefore, DC planners need to be very conscious when looking for cities for building new DCs.

Data Centers use diesel generators for backup power. This can have a huge impact on the health of the residents living in the city. Town of Quincy in Washington is home to huge data centers of Microsoft, Yahoo and Intuit. Many other companies are in line to setup their data centers in this small town in central Washington State. Although the economic development officials are happy to setup more businesses in the town, ecology department and urban planners are worried about the growth of data centers in their town. (Quincy 2011)

7. Application of the model

Companies use different criteria for their selection of sites. Decisions for a new site can be based purely on financial considerations, or avoiding disaster. Some companies want their DCs near to their corporate offices, others want their DCs spread out due to business continuity reasons.

We studied different factors and talked to industry experts for developing this model. In this paper, we are considering that a company needs to set up a new DC. Company in question might be a bank, a hospital or an IT services provider. We are developing a model which their executive team can use for selection of a city among different alternatives for choosing a DC site.

A group of experts were selected to provide the following data required to execute the mode: (1) The constant-sum values representing comparative judgement on each pair of criterion to determine the relative priority of the seven criteria. (2) The constant-sum values representing comparative judgements on the set of factors associated with each criteria. (3) A rank between 0 and 100 representing his/her judgement on the relative desirability of each factor against each alternative (the mean values were calculated among the relative values given by each expert to represent the group decision). The following is a brief profile for each of the respondents (experts):

Expert 1	=	IT DC manager – responsible for operations of DC.
Expert 2	=	facilities planner – responsible for design and construction of buildings.
Expert 3	=	IT manager – responsible for infrastructure which includes all IT equipment.
Expert 4	=	finance analyst – responsible for the net present value (NPV) and return on investment (ROI) analysis and budgeting, etc.
Expert 5	=	electrical engineer – responsible for cooling and power issues in a DC.

The following four steps were followed to run the calculations based on the HDM above and the data collected from the experts:

Step 1: Determine the weight of each criterion by determining the relative priority of that criterion with respect to the objective. As mentioned earlier, the constant-sum values representing comparative judgement on each pair of criteria were obtained from each expert to determine the relative priority of the four criteria. Table 1 shows an example of constant-sum values for all criteria provided by one expert.

Table 1 Example of criteria comparative analysis provided by one expert.

Priority of criterion-1		Priority of criterion-2
C1: geographical factors	55	45	C2: financial factors
C1: geographical factors	70	30	C3: political factors
C1: geographical factors	65	35	C4: social factors
C2: financial factors	55	45	C3: political factors
C2: financial factors	70	30	C4: social factors
C3: political factors	50	50	C4: social factors

Using PCM software (PSU 1991), the relative weight of each of the four criteria was calculated based on the data provided by each expert. The final weight for each criterion was found by calculating the mean weight from all experts. Table 2 shows the final weights for each criterion in this model.

Table 2 The relative weight of each criterion.

Decision criteria	C1	C2	C3	C4	Σ
Relative weight	0.28	0.33	0.19	0.19	1

Step 2: Determine the relative impact (weight) of each factor under each criterion. The constant-sum values representing comparative judgements on the set of factors associated with each criterion were obtained from all experts. The relative importance of factors with respect to their criterion was calculated by following the same approach as step 1 above. The final weights representing the group mean for the normalised relative importance of factors under each criterion are given in Table 3.

Table 3 Relative weights of factors under each criterion.

C1: geographical factors	F11	F12	F13	F14	F15	Σ
Relative weights	0.23	0.19	0.24	0.21	0.13	1
C2: financial factors	F21	F22	F23	Σ
Relative weights	0.31	0.21	0.49	1
C3: political factors	F31	F32	Σ
Relative weights	0.66	0.34	1
C4: social factors	F41	F42	Σ
Relative weights	0.56	0.44	1

Step 3: Determine the rank or desirability of each alternative considering all factors under all criteria. Each expert provided a rank between 0 and 100 representing his/her judgement on the relative desirability of each alternative when associated with each factor (the mean values were calculated among the relative values given by each expert to represent the group decision). Table 4 shows the mean desirability for each alternative against each factor.

Table 4 Ranking of alternatives against each factor.

		Desirability values (0–100)
Criterion	Factors	Alt1: Quincy, Washington	Alt2: Sacramento, California	Alt3: Charlotte, North Carolina	Alt4: Dalles, Oregon
C1: geographical factors
	F11: disaster avoidance	83	66	79	82
	F12: transport availability/accessibility	66	85	85	71
	F13: telecom network availability	66	84	86	79
	F14: power availability	87	73	83	83
	F15: water availability	77	65	74	81
C2: financial factors
	F21: land cost	85	65	70	85
	F22: building construction cost	81	74	72	86
	F23: variable costs – electricity cost, property tax	87	68	76	83
C3: political factors
	F31: tax structure, incentives and subsidies	90	73	74	92
	F32: jobs creation	74	77	75	67
C4: social factors
	F41: safety and security, crime	85	69	73	85
	F42: laws related to urban planning	84	62	63	83

Step 4: Calculate the value of each alternative using the output data from steps 1, 2 and 3. This is done by applying the summation of the mean weight of each criterion times the mean weight of each related factor times the desirability value of the alternative for that factor. This is done for each alternative. Table 5 shows the computation of the final value for the first alternative as an example. The following formula summarises the mathematical formula used to obtain the final value for each alternative. As indicated in Table 5, the final value of the first alternative is Alt1 = 81.72. The same calculations were made for the rest of the alternatives (Alt2 = 70.5, Alt3 = 74.56 and Alt4 = 82.12).

where A _v represents alternative final value; Cw _i , weight of criterion i (i = 1 − y); Fw _ij , weight of factor j in Ci (j = 1 − k); D _ij , alternative ranking for factor j in C i; y, number of criteria in the model; k, number of factors under C i.

Table 5 Calculations for alternative-1 (Quincy, Washington) final value.

Criterion	Factors	Alternative desirability value	Alternative value
C1: geographical factors (0.28)	F11 (0.23)	83	5.35
	F12 (0.19)	66	3.51
	F13 (0.24)	66	4.44
	F14 (0.21)	87	5.12
	F15 (0.13)	77	2.80
C2: financial factors (0.33)	F21 (0.31)	85	8.70
	F22 (0.21)	81	5.61
	F23 (0.49)	87	14.07
C3: political factors (0.19)	F31 (0.66)	90	11.29
	F32 (0.34)	74	4.78
C4: social factors (0.19)	F41 (0.56)	85	9.04
	F42 (0.44)	84	7.02
		A v	81.72

8. Discussion

The results generated by executing the model indicate that alternative-1 (Quincy, Washington) scores 81.72% of the ideal DC location for the company in this mode. The same applies for alternative-2 (Sacramento, California), alternative-3 (Charlotte, North Carolina) and alternative-4 (Dalles, Oregon) with scores of 70.5, 74.56 and 82.12%, respectively.

The results suggest that alternative-4 (Dalles, Oregon) has the highest score among all alternatives followed by Quincy and Washington with a score of 81.72%. This indicates that selecting Dalles as a location for the DC is 1.005 times as preferable as the second best alternative Quincy. It is obvious that the highest scoring alternative has a very negligible advantage over the second which does not provide the decision maker with clear evidence about the advantage of the best alternative. In this case, we suggest implementing one of the following two alternatives in order to fine-tune the model and provide precise values for competing alternatives to the decision maker:

1.	Rerun the entire model against the best two alternatives only. Dropping all other alternatives from the model will put the competing alternatives in direct competition and allow experts to provide more accurate data at each step of the model.
2.	Instead of using the ranking technique in step 3, the constant-sum method can be used again to determine relative weight of each alternative with respect to each factor. This method will provide a direct comparative analysis among alternatives and will allow experts and decision makers to provide a more precise data.

HDM can be used for a large number of criteria and factors keeping the same structure. Changing the number of experts and alternatives does not alter the structure of the model. Therefore, companies can adopt this model for any size of the decision model.

Robustness is a characteristic of the model and not of the data. As long as a model is not sensitive to the errors in the data, it can be categorised as a robust model. In this type of scenario, when we have many variables to consider for a decision, simple mathematical models are not sufficient. This model has taken inputs from experts from different domains of an IT company who understand the challenges and issues in the operations of DCs. As a part of the validation we shared the modelling approach and our results with experts from regional organisations in the energy sector. They have confirmed our modelling approach and our results.

HDM with pairwise comparisons has been extensively used in various areas such as technology selection, software tool selection, manufacturing, automobiles, electronics and project management to study different goals using a combination of criteria and alternatives. When making investment decisions, organisations need to consider all data. Intuition-based decisions can be very costly to the company. Mathematical based decision models can help the management in making the right investment decisions in the best interests of the organisation (Patanakul et al. 2007).

9. Conclusions

DCs are at the core of any business enterprise operations. DC industry is going through a transformation and getting matured. Performance is no longer the sole criteria for defining the success of DCs. Energy efficiency, resources optimisation and business continuity are also as important as the technology of IT equipment itself. With so much at stake, companies need to plan accordingly in order to deliver the best business value to the enterprise. Companies will find this model valuable as this can be adopted for decision-making without major investment.