If you bill it, they will pay: Energy consumption in the cloud will be irrelevant until directly billed for Abram Hindle Department of Computing Science University of Alberta, Edmonton, Canada Email: hindle1@ualberta.ca Abstract—Don’t leave the lights on! One reason we take energy Even office computers are blamed for city-sized levels of consumption seriously is because we are directly billed for it. If energy consumption [2], [3]. Yet mobile energy consumption one leaves the heat on high over night the effect is noticeable is not significant environmental sustainability issue [4]. The on the next bill. Yet as granular as cloud computing billing can be in terms of resources and quality of service, we have limited chemicals used in batteries are a concern but the joules used motivation to investigate energy consumption in the cloud because on mobile devices pale in the comparison to the joules of an cloud customers cannot necessarily realize savings. Proxies for idle server in a datacenter being cooled by air conditioning. quality of service such as lower performance CPU allocations Thus server-side software energy consumption is one of the can be used but at no point does a user see a bill listing biggest sustainability concerns of software. energy consumption. Furthermore the difficulty in billing energy consumption of virtualized services is non-trivial and indirect. Cloud providers face tough decisions regarding energy When many VMs share the same host, attribution of energy consumption as elasticity and availability are buffered by the consumption becomes difficult. When many hosts are in the same high cost of idle machines in the datacenter. High utilization is datacenter attribution of cooling costs become difficult as well. a goal, achieved by slicing servers up into VMs and containers. Thus due to the direct and indirect costs of running a cloud, Yet high utilization has many other costs, heat, maintenance, and the sharing of resourcing pricing cloud energy consumption is difficult and typically not done. We argue that until energy energy, poorer per VM performance, are just a few. Heat consumption of hosted computers, VMs, and cloud services is requires further cooling thus consuming even more energy. pushed down from the cloud provider to the cloud consumer, While it is easy to slice up the resources of a server and datacenters will continue to consume massive amounts of energy add it to a cloud, it is hard to attribute energy consumption to provide software services. When cloud end-users have to pay to each particular VM, container, or service on a cloud for energy consumption they will consider optimizing energy consumption. Once energy consumption in the cloud is a bill line server, although there is much research in the area [5], [6]. item, energy consumption will become a first class performance Performance counters are the simplest method of performance non-functional requirement of software. monitoring yet they do not work for counting energy consumed by cloud services. Cloud services therefore will need to I. I NTRODUCTION leverage performance counters to suggest or estimate the actual We propose a lightning talk to be delivered within 5 to 10 energy consumption in the cloud. Furthermore knowledge minutes at RE4SUSY 2018. The lightning talk will cover the of the global data-center will be required to understand the topic of lack of demand for software sustainability features externalities of the energy consumption of computation. until cloud services start billing for energy consumption. We argue that nothing will motivate cloud end-users to During a discussion with some unnamed industry partners optimize for such an invisible externality until it becomes we asked about energy consumption, they responded that a reality on their cloud services bill! Until the motivation they were concerned, but until they could put a price on for sustainability can be expressed monetarily on a bill for energy consumption of their software systems, they were services, there will be few requirements communicated from going to ignore it. Why bother to optimize or even plan customers to address software energy consumption of their for energy consumption when you had immediate pressing cloud services. Given the current trend in cloud billing where concerns from requirements and non-functional requirement giants such as Amazon rate CPU time and even engage in CPU alike including features, stability, reliability, and cost of cloud credit systems with fine granularity there is likely pressure services. Typically cost of cloud services already includes to bundle the cost of the externalities of a data-center into energy consumption and it is not a line item—thus optimizing the cloud costs as well. The closer the customer can get to for energy consumption does not have a clearly observable paying for exactly what they using the easier both parties can effect on the cost. Without insight into, feedback about, or the plan and expand. Thus we argue there is pressure on the part cost pressure of energy consumption, it will not be optimized of cloud providers such as Amazon, Microsoft, and Google, for or even requested for from customers. to ensure that they can push all of their costs to the end- Energy consumption of software systems and services is a user proportionally. The value in not direct billing is that the grave concern as data-centers now account for a significant customer assumes it is accounted for, but what the customer portion of the worldwide energy consumption (2-3% [1]). cannot see or measure, they will not optimize for. A customer Copyright held by the author(s). who uses non-optimal instance VMs has no incentive to avoid concern and competitive for cloud services. We believe that the this behaviour—for example low memory VMs coupled high advent of itemized billing for energy and other sustainability numbers of CPU cores waste memory as they deny the unused concerns will prompt cloud-services end-users to care and memory to other VM instances who cannot be scheduled to optimize for energy consumption concerns. Conversely, as that machine until the cores are freed up. Knuth stated, “pre-mature optimization is the root of all We argue that the incentives for sustainability should be, and evil” [13], cloud-service end-users will not optimize for what eventually will be, pushed down to the cloud-service end-user. they cannot measure or see. Until that point, few will prematurely optimize for energy, and Cloud providers will have to convince users to adopt a sustainability when they can directly optimize for other costs. billing scheme based on energy consumption estimation as it cannot be measured directly. Direct measurement is not a II. D IRE N EED large barrier as the governments of world propose taxation Energy consumption can come at a high cost, Song et and incentivization strategies to address sustainability issues al. [7] show that 40% of the energy cost of a server is in and many unmeasurable externalities. It is unlikely that pri- cooling that server, which means for a small server running at vate service providers cannot enact what governments have 300W for 5 years, not only would nearly $1000 USD be spent successfully deployed [10]. powering the server—given the US average rate of 0.0733 per Cloud providers should bill customers directly in order to kWh [8] (.3kW · 24h · 365d · 5y · $0.0733) —but $600 USD motivate requirements of sustainability and foster further inno- would be spent cooling that same server. The externalities of vation in software requirements and software design relevant data center cooling are such a concern that Microsoft has to software energy consumption and sustainability. experimented/marketed with underwater datacenters [9] that use the ocean as a giant heat-sink. R EFERENCES As data-centers are prone to deploy near where people [1] T. Bawden, “Global warming: Data centres to consume three times as actually live—in order to reduce latency—they often prone much energy in next decade, experts warn,” in Independent (Newspaper). Independent, 2016, https://www.independent.co.uk/environment/global- to rely on energy produced in the local area—for instance warming-data-centres-to-consume-three-times-as-much-energy-in- in Edmonton, Alberta much energy is provided by coal fired next-decade-experts-warn-a6830086.html. power plants which have signficant environmental and health [2] US Department of Energy, “How is electricity used in u.s. homes?” https://www.eia.gov/tools/faqs/faq.php?id=96&t=3, Feb 2018. impacts. Often this energy has significant carbon emissions [3] E. Star, “Energy star: The simple choice for energy efficiency,” 2016. and thus data-centers energy will also be taxed accordingly [Online]. Available: https://www.energystar.gov with carbon-taxes. We argue that much like carbon-taxes, end- [4] A. Hindle, “Green software engineering: The curse of methodology,” in 23rd IEEE International Conference on Software Analysis, user billing of software energy consumption will promote Evolution, and Reengineering (SANER 2016) FOSE Track: Leaders of reduced energy consumption or at least sustainable energy Tomorrow: Future Of Software Engineering, 2016, inproceedings, consumption much like economic models of carbon tax predict pp. 529–540. [Online]. Available: http://softwareprocess.ca/pubs/ hindle2016SANERFOSE-green-software-engineering.pdf will happen with emissions and emission producers [10]. [5] M. Kurpicz, A.-C. Orgerie, and A. Sobe, “How much does a VM cost? Energy-proportional Accounting in VM-based Environments,” in III. P OSSIBLE E FFECTS PDP: Euromicro International Conference on Parallel, Distributed, If end-user billing of software energy consumption of cloud and Network-Based Processing, Heraklion, Greece, Feb. 2016, p. 8. services became possible then those hosting services would [Online]. Available: https://hal.inria.fr/hal-01276913 [6] R. Shea, H. Wang, and J. Liu, “Power consumption of virtual machines have another dimension of performance to optimize for. Many with network transactions: Measurement and improvements,” in INFO- other researchers have argued that users might be willing to COM, 2014 Proceedings IEEE. IEEE, 2014, pp. 1051–1059. trade off one concern to improve a performance concern such [7] Z. Song, X. Zhang, and C. Eriksson, “Data center energy and cost saving evaluation,” Energy Procedia, vol. 75, pp. 1255 – 1260, as network bandwidth or energy consumption [11], [12].For in- 2015. [Online]. Available: http://www.sciencedirect.com/science/article/ stance Zhang et al. [12] propose ranking and comparing appli- pii/S1876610215009467 cations based on energy consumption a system they call soft- [8] M. Rareshide, “Power in the data center and its cost across the u.s.” https://info.siteselectiongroup.com/blog/power-in-the-data-center- ware application energy consumption ratings (SAECR)/Green and-its-costs-across-the-united-states (Accessed 2018-07-20), 10 2017. Star—like Energy Star [3]). Thus system operators might [9] R. Cellan-Jones, “Microsoft sinks data centre off orkney,” https://www. care about the performance rankings of server-side software bbc.com/news/technology-44368813, June 2018. [10] X. Liu, Y. Leung, Y. Xu, and L. C. W. Yung, “The effect of carbon if they were being billed for sustainability concerns such as tax on carbon emission abatement and gdp: a case study,” Journal of energy consumption. If system operators are choosing software Geographical Systems, vol. 19, no. 4, pp. 399–414, Oct 2017. [Online]. packages and services for their sustainability footprint then Available: https://doi.org/10.1007/s10109-017-0254-1 [11] R. Saborido, F. Khomh, A. Hindle, and E. Alba, “An app there will be significant pressure on software developers to performance optimization advisor for mobile device app marketplaces,” address energy consumption as a first class non-functional Sustainable Computing, pp. 1–18, May 2018. [Online]. Available: http: requirement in their software systems. //softwareprocess.ca/pubs/saborido2018SUSCOM-app-optimization.pdf [12] Z. Chenlei, A. Hindle, , and D. M. German, “The impact IV. C ONCLUSION of user choice on energy consumption,” IEEE Software, pp. 69–75, 2014. [Online]. Available: http://softwareprocess.ca/pubs/ Thus we have argued for the need for software energy zhang2014IEEESoftware-user-choice.pdf consumption billing in cloud computing services—we believe [13] D. E. Knuth, “Structured programming with go to statements,” ACM Computing Surveys (CSUR), vol. 6, no. 4, pp. 261–301, 1974. it is inevitable as price and quality of service is a constant