Assessing anthropogenic heat flux of public cloud data centers: current and future trends

Architecture of public clouds

Major cloud service providers (CSP) (i.e., Google, Microsoft Azure and Amazon) build their infrastructure in a way that would make their services reachable from almost any place on Earth, as long as there is an internet connection. That is, Data centers (DCs) are distributed among distinct regions, which are located in different cities, countries, and continents. Currently,² DCs of public CSPs are distributed among 65 locations around the world. As clarified in Fig. 2, each region consists of at least one availability zone, and each zone consists of at least one DC, with its independent energy sources, cooling systems, and communications. Such formation allows virtual networking and increases the availability to reach 99.99% in some cases (Uzaman et al., 2019). Table 1 presents the number of regions and zones currently available from each of the top three CSPs. As it can be noted in the table, the total number of regions is 106, with average number of zones per region ≈ 3. Taking the average number of DCs per zone ≈ 3 (as publicly published on the considered CSPs websites), the total number of DCs that belong to these CSPs ≈ 933 DCs.

Power consumption of public cloud data centers

Reports indicated that cloud DCs are responsible for more than 2% of the US total electricity usage, and 1–1.5% of global electricity usage (Andrae & Edler, 2015). However, different studies revealed large variations of the energy consumption of the DC industry as illustrated by the results of five different studies (Table 2). The average value of the results of these studies is 208.4 billion kWh/year, with a standard deviation of 116.5 billion kWh/year. Even though this average value is derived from estimates of annual global energy consumption in different periods of time (2010–2020), during which the total number of DCs and the DC technology have changed, it is very close to the result (205 billion kWh/year) recently reported by Masanet et al. (2020) for the global energy consumption of DCs in the year 2018. Additionally, if the projected global DC energy consumption in the year 2020 is estimated from the results provided by Ascierto et al. (2015) and Andrae & Edler (2015), the value of (270 ± 130) billion kWh/year is also consistent with the above mentioned average. The large variations of the reported results in the individual studies is attributed to adopting different numbers of DCs, and different methodologies.

As our intention in this article is to determine the AHF produced by the DCs of the three public CSPs (a total of 933 DCs), knowledge of the estimated energy consumption and efficiency of a single DC is required. Analysis reported by Kaplan, Forrest & Kindler (2008) estimated that a public cloud DC consumes as much energy as 25,000 households. Bearing in mind that the average monthly energy consumption of a household at the time of that study was 339 kWh (Firth et al., 2008), the average energy consumption of one public DC is then estimated by 101.7 × 10⁶ kWh/year. Therefore, the total energy consumption of DCs belonging to the three CSPs is 101.7 ×10⁶ × 933 = 94.9 billion kWh/year. This accounts for 46% of the global average of 208.4 billion kWh/year. Recent reports estimated the global market share³ of Amazon, Microsoft Azure, and google as 32%, 19%, and 7% of the global CSPs market, respectively. According to these market shares, the three CSPs consume 58% of the global average energy consumption (120.9 billion kWh/year). Consequently, our best estimate of the uncertainty in Kaplan’s estimate of the energy consumption of a single DC is obtained from the difference between these two results ([120.9–94.9]/94.9) × 100% = 27%). According to Kaplan’s estimates, the average power consumption of a public DC ≈ 11.6 MW (± 3.132 MW, bearing in mind that a year is about 8760 h). In fact, this average is in good agreement with the average power consumption of 10.17 MW per DC, during the year 2018, derived from Google’s report Google (2019). In the forthcoming analysis, we will adopt Kaplan’s estimates of the energy consumption by a single DC.

Power distribution and efficiency in public data centers

The data center industry uses the Power Usage Effectiveness (PUE) metric to measure a DC efficiency (Gadgil & Rahn, 2019). The PUE of a DC is the ratio of the data centre total energy consumption P_Total to information technology equipment energy consumption P_IT, calculated, measured or assessed across the same period (ISO/IEC 30134-2:2016, 2018). PUE is calculated using Eq. (1) (Gadgil & Rahn, 2019). The distribution of power consumption in a typical DC was described by Rasmussen (2006), and a schematic representation is depicted in Fig. 3. (1)${\displaystyle PUE=\frac{P_{Total}}{P_{IT}}.}$

Two experiments were conducted by Marcinichen, Olivier & Thome (2012) to determine the efficiency of a DC consuming a total power of 900.8 W and 1991 W. The power consumed to perform computational tasks and maintain the IT equipment was found to be 428.1 W and 890 W, respectively. These results indicated a PUE value of ≈ 2.10 and 2.24, respectively. These PUE values are in good agreement with the experiments conducted two years later by Zhao et al. (2014), which resulted in PUE value ≈ 2 for the studied DC. These measures are higher than typical PUE values recently averaged by 1.67 (Ponnusamy, Sharma & Lee, 2021). Additionally, those studies did not use the ISO 30134-2 definition of PUE, which must be measured for one year and where the energy meters are located. Nonetheless, more efficient DCs may waste only around 10% of the consumed energy (Wahlroos et al., 2018) resulting in a lower PUE value, and thus a higher efficiency.

Estimations regarding waste energy in different DCs can fluctuate depending on the fluctuation of workloads, the precise gross floor space (GFS) (Boehme, Berger & Massier, 2015), and the number of CPUs/machines. Thus, public CSPs tend to provide average PUE values for all their infrastructure as an indicator of the average efficiency of all their equipment. Google official reports (Google, 2020) claimed an average PUE value of 1.10 during the past 12 months (as of Nov.2020). On the other hand, Microsoft average PUE was reported to be 1.25 (Diekmann, 2020), whereas the most recent report about Amazon cloud PUE indicated an average value of 1.14 (Mah, 2016). The fraction of energy waste (= 1–1/PUE) indicate roughly 9.1%, 20.0%, and 12.3% energy waste by the three CSPs facilities, respectively.

Modeling DCs and their required cooling systems are mathematically presented in details in Uchechukwu, Li & Shen (2014). Specifically, detailed criteria related to cooling technologies and energy consumption of DCs are analyzed in Rasmussen (2005). Generally, the temperature of the air in the servers should be kept in the range of 15–20 °C (Ashrae, 2015). However, the cooling process transfers heat into the surroundings of the center in the form of AHF. Typically, cooling systems require large supply of energy (i.e., electrical, mechanical, etc.), resulting in significant amounts of emitted heat flux.

The heat flux resulting from different types of DC processors was reported in the literature, such as in Patel (2003); Marcinichen, Olivier & Thome (2012); Trutassanawin et al. (2006); Samadiani, Joshi & Mistree (2008). However, the total waste heat generated by those DCs was rarely investigated due to the variety of DCs’ architecture and cooling approaches. Jain et al. (2013), for instance, described how cooling systems in public cloud’s DCs run 365 days a year for full 24 h per day, and how to calculate the waste energy, performance per watt, the greenness of the cloud system, among other criteria of the clouds. They also described major solutions proposed in the literature to decrease the energy consumption of such systems. On the other hand, Huang et al. (2020) analysis revealed that the quality of waste heat in two-phase cooled systems is the highest among the three types of the analyzed cooling systems, due to the higher heat transfer efficiencies. That is, using a two-phase cooling system allows for better waste heat in terms of quality, which leads to higher revenue in case the waste heat was harvested. For air-cooled systems, the quality of waste heat is low, and thus heat pumps, e.g., Bamigbetan et al. (2019), may be added to the system if the waste heat is to be harvested.

Global AHF estimations

The average global warming is reported to be 1 °C above pre-industrial levels (Delmotte et al., 2018), and is expected to go up in the foreseeable future (Saklani & Khurana, 2019) to reach a minimum of 4 °C above pre-industrial levels by 2100 if the current conditions remain Stott et al. (2006). Several workers around the world investigated this phenomenon and related parameters, including AHF emissions, and proposed potential solutions for the global warming problem. For example, Stewart & Oke (2012) proposed a classification of Local Climate Zones (LCZs) as a model for a comprehensive climate-based classification of urban and rural sites, and suggested 17 different classes of LCZs depending on different parameters, one of which is anthropogenic heat emissions. On a local-scale level, Gabey, Grimmond & Capel-Timms (2019) compared simple and detailed AHF models in London at 500 m² spatial resolution (i.e., they divided the city area into blocks, each with 500 m² area). The authors used population data, national and local energy statistics, and traffic on the roads network data in their analytical study. Accordingly, they measured high levels of AHF in the city center and main roads (roughly 10% of the city area), which accounted for 30–40% of total energy consumption of the city.

On the continental and global scales, Flanner (2009) investigated the AHF over three regions, namely the continental United States, Europe and East Asia, and demonstrated that AHF may indeed contribute to the global warming, and hence should be included in climate simulations. Here, Flanner analyzed the AHF effect in the period of 2005–2100 using historical and real-life measurements. The 2005 results revealed averaged AHF densities of 0.39, 0.68, and 0.22 W m⁻², respectively, which were predicted to increase in 2040 up to 0.59, 0.89, and 0.76 W m⁻², respectively. The results of the study also predicted an annual-mean warming of 0.4–0.9 °C over large industrialized regions in 2100 as a result of global AHF emissions. Finally, the author reported Global-mean AHF densities of 0.028, 0.059, and 0.19 W m⁻² in the years 2005, 2040, and 2100, respectively. Recently, however, global AHF densities in 2020 were reported as 0.12 W m⁻² (Lu et al., 2017), and 0.15 W m⁻² (Jin, Wang & Wang, 2020), significantly higher than the value that can be estimated by extrapolating Flanner’s results.

Solutions for utilizing AHF

Bamigbetan et al. (2019) proposed the employment of a heat pump to transform the waste heat into a usable form. Similarly, Deymi-Dashtebayaz & Valipour-Namanlo (2019) investigated the feasibility of using the air source heat pump (ASHP), as a waste heat recovery system operating with different refrigerants, in a 54-racks data center. The study indicated a consequent 35,000 m³/year natural gas saving, 20.8 MWh/year electrical energy saving, 121 tons/year CO₂ emission reduction, and 25,000$/year cost saving.

Sahana, Bose & Sarddar (2018) proposed positioning a group of sensors into cloud DCs to monitor and minimize cooling efforts, and hence contribute to the energy efficiency of the cloud. Yu et al. (2013) proposed an algorithm to predict the temperature of cloud DCs, in order to facilitate the deployment of energy efficient solutions and predicting energy consumption. Sobhanayak & Turuk (2019) proposed a thermal-aware task allocation and scheduling algorithm for cloud DCs, aiming to minimize power consumption that is involved in cooling processes.

Pärssinen et al. (2019) provided an economic investment assessment of DC waste heat utilization in three different size cases, namely small, medium and large. Accordingly, they concluded that medium and large cases (e.g., DCs of public CSPs) can be considered profitable business opportunities and should result in a decision to invest in a waste heat reuse solution. For the biggest barriers for utilizing waste heat are the low quality of waste heat and high investment costs, Wahlroos et al. (2018) proposed A systematic 8-step change process to ensure success in changing the priority of AHF utilization in the DC and district heating market. Similarly, Alhamwi et al. (2018) showed how investing in local electricity storage and on-site renewable power generation solutions can significantly reduce the total DC system costs.

Current AHF emissions by public CSPs

The results in Tables 3 and 4 indicate that:

1. With a total waste heat at a rate of 1,211.04 MW, Microsoft Azure contributes more than Amazon and Google collectively, whose contributions are 299.628 and 209.844 MW, respectively. The contributions of the individual CSPs relative to the cumulative waste heat by the three CSPs are, respectively, 70.4%, 17.4% and 12.2%. Interestingly, these values could not be expected according to the market shares discussed in ‘Power consumption of public cloud data centers’.

2. Azure and Google’s highest waste heat rates are in the continents of N.America, Asia, and Europe, respectively. While the highest waste heat rates emitted by Amazon are in Asia, N.America, and Europe, respectively. Figure 4 illustrates these results.

3. An average of 2.88% out of total continental AHF density is contributed by DCs that belong to the studied CSPs (i.e., Azure, Amazon and Google). This percentage can be divided using the ratios presented in point 1 above.

4. Although Europe receives less total waste heat than N.America and Asia, as illustrated in Fig. 5, it receives the highest AHF density due to its relatively small area compared to N.America and Asia. That is, Europe atmosphere receives more than 25% out of the total waste heat emissions by all DCs under consideration. On the other hand, lowest AHF densities emitted by public DCs can be found in S.America, Asia and Africa, respectively. This is due to their large areas that include small number of public DCs, and/or low SNSR values. These results are illustrated in Fig. 6.

Future trends of AHF emissions of public clouds DCs

Based on a top-down energy inventory approach and up-to-date data, Lu et al. (2017) provided globally averaged AHF estimations up until the year 2100, considering high, moderate, and low AHF emissions scenarios. On the other hand, the percentage of averaged current continental AHF, emitted by the studied CSPs, relative to averaged current continental AHF of all activities, was calculated in Table 4 and equals 2.88%. Consequently, we deploy this percentage into Lu et al.’s estimations, to deduce future trends of AHF of, specifically, DCs of the studied public CSPs. Similarly, Fig. 7 demonstrates our estimations concerning high, moderate, and low growth in the Data Center Industry, until the year 2100.

As can be noted in the figure, average continental AHF, emitted by public clouds DCs is expected to reach a value between 0.005 and 0.008 (± 27%) Wm⁻². Such AHF value represents a percentage of 2.63–4.21% of the mean global AHF density estimated by Flanner.