Estimates of annual data center electricity usage vary from 200 terawatt hours (TWh) [43] to 500 TWh [11]. The lower of these figures would suggest that data centers consume 1% of global electricity [43], but this could be significantly higher. One study suggests that global data center energy usage was 270 TWh in 2012 [66]. Another study estimates that 104 TWh will be used by European Union data centers in 2020, which makes a global total of 200 TWh unlikely [9].
This uncertainty also extends to efficiency estimations. As of 2018, IT (Information Technology) workloads have grown significantly compared to 2010 - ×6 more compute instances, ×10 more network traffic and ×25 more storage in 2018, yet it is reported that data center energy usage has only grown by 6% over that time [49]. Some reports support this by showing that average Power Usage Effectiveness (PUE) is improving [9], but industry survey data suggests average PUE values have plateaued [45]. Future energy usage is uncertain: efficiency improvements may be “frozen” [61] and some scenarios suggest data center energy usage could double by 2030 [7].
Whether data center energy usage grows modestly or increases significantly, even with the increasing use of renewables in the technology sector [44] data center emissions and other associated environmental impacts still need to be accounted for.
In the past, IT was run in-house (also known as “on-premise”). IT teams would buy physical servers, disks and network devices from vendors such as Dell, Seagate and Cisco, then install them into data centers. These data centers might be built and operated by the company itself, or space would be rented (“co-located”) in large scale facilities, such as those run by Digital Realty or Equinix. The company would pay for the space required to deploy the number of servers they needed, install internet access and purchase power, typically over-provisioning their equipment to ensure they had spare capacity.
IT applications running on physical hardware have a known (or knowable) footprint. The equipment is self-contained and can be traced to a manufacturer so the embodied cost of components can be calculated. Data center characteristics such as power and cooling levels can be monitored. Emissions factors for the electricity mix can be determined. As such, it is possible to calculate the environmental footprint of a deployment.
Guidelines exist for creating energy efficient data centers [38] and metrics such as Power Usage Effectiveness (PUE) can be calculated [40].
PUE is a widely used metric and often cited to show progress in data center efficiency. For example, Google publishes quarterly and trailing 12-month PUE values going back to 2008 for their global fleet of 15 data centers [29]; the latest Google Q1 2020 fleet wide PUE is 1.09. However, PUE has been criticised when used as a measure of efficiency because it only considers energy. PUE can decrease when IT load increases even though efficiency may not have improved [16]. It has also been shown to correlate poorly with carbon emissions [47, 48] and should not be the only metric tracked [70].
Water Usage Effectiveness as a site based metric (WUE), combined with its complementary source based metric (WUEsource) [57], are important environmental indicators because of the large volumes of water that data centers require, projected to be 660 billion litres for US data centers in 2020 [60]. Most of this water is used in electricity generation, which is why the WUEsource metric includes external water intensity factors, not just the operational water usage at a point in time [57]. Although moving to renewable sources of electricity generation helps reduce WUEsource because wind and solar energy have low water footprints [64], less than a third of data center operators track any water metrics [36]. Facebook is one of the few companies who report both PUE and WUE figures publicly [21]. Other metrics such as Renewable Energy Factor (REF) [41] and Energy Reuse Factor (ERF) [39] exist as international standards but are difficult to find in public disclosures.
Understanding when to refresh hardware is another element to consider. In a survey of European data centers, IT equipment older than 5 years was shown to consume 66% of energy despite only representing 7% of capacity [11]. However, replacing equipment less than 4.5 years old may cost more in hardware than is saved on energy efficiency [11]. This highlights the importance of lifecycle analysis because hardware refresh rates and overall utilisation impact the environmental footprint of a data center, potentially offering more energy savings than decreasing PUE [10].
With the availability of cloud computing services from vendors such as Amazon Web Services (AWS), Google Cloud Platform (GCP) and Microsoft Azure, workloads are increasingly being deployed into public cloud services [23]. This trend is demonstrated by the growth of the global cloud computing market over the last decade. From just under $6bn in 2008, as of 2019 it had reached $208bn and is projected to grow to $236bn in 2020 [23]. Server equipment purchases are growing at 3% per year, almost entirely attributable to “hyperscale” cloud vendors [60]. Estimates suggest 40% of servers will be in hyperscale data centers in 2020 [60].
This move to the cloud has made it much more difficult to estimate associated emissions. Public cloud vendor customers purchase virtual services so it is difficult to know what underlying physical resources are used because they have been abstracted by complex software or platform layers. Customers migrating to the cloud must also ensure their cloud architecture is equivalent to their on-premise hardware deployments in terms of availability and redundancy to ensure that comparisons are accurate. Cloud vendor customers have no insight into the energy usage of the services they buy, and often do not even know how many physical servers their applications are running on. Instead, they pay for precise usage such as CPU time, allocated memory or execution time. In theory the price should include the full costs of components like power and disks, but the number is not transparent. Much is hidden behind opaque cloud vendor pricing. Some vendors have used marketing efforts to explain why public cloud is “greener” than on-premise [1, 51] but do not provide specific, detailed numbers behind their claims. Models such as CLEER [47, 48] can make assumptions, but there are so many variables that their accuracy is questionable, particularly across use cases and as the model ages.
With an increased public awareness of environmental issues [69] and more businesses being covered by mandatory reporting [20], cloud vendor customers should expect to be able to calculate the environmental footprint of their IT environment just as if they were running it on-premise. This has been possible historically: the GHG Protocol reporting guidance assumes a range of measures such as server count, data center PUE and capacity are available [27]. However, the current approach by public cloud computing vendors makes it difficult to obtain that information.
Several studies [13,14,15, 17, 24, 25] have proposed new approaches where workloads can be dynamically moved based on various “follow-the-renewables” criteria [15] about the underlying data center e.g. regional wind power availability. These techniques can work well if the system has access to data to make the right decisions e.g. emissions factors related to the data center energy input, external temperatures related to cooling requirements, wind speeds in the relevant geographies for each data center, etc. If the customer is running their own data center or is deployed in a co-location facility, then they should be able to get access to this information. However, this is not possible with public cloud. As this paper will discuss, the major cloud providers do not release most of these underlying data, and if they report anything it is usually only in aggregate, not real-time.
This paper develops a framework for understanding the boundaries of a public cloud computing environment, then uses that framework to evaluate whether the Greenhouse Gas (GHG) Protocol is suitable for calculating emissions from cloud workloads. It also considers what cloud vendors have done and should do in the future to allow customers of public cloud to calculate their own environmental footprint.