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

Paper #

Leiserson, C.E., Thompson, N.C., Emer, J.S., Kuszmaul, B.C., et al. (2020) There’s plenty of room at the Top: Baniata H, Mahmood S, Kertesz A. 2021. Assessing anthropogenic heat flux of public cloud data centers: current and future trends. PeerJ Computer Science 7:e478 https://doi.org/10.7717/peerj-cs.478

Abstract #

Global average temperature had been significantly increasing during the past century, mainly due to the growing rates of greenhouse gas (GHG) emissions, leading to a global warming problem. Many research works indicated other causes of this problem, such as the anthropogenic heat flux (AHF). Cloud computing (CC) data centers (DCs), for example, perform massive computational tasks for end users, leading to emit huge amounts of waste heat towards the surrounding (local) atmosphere in the form of AHF. Out of the total power consumption of a public cloud DC, nearly 10% is wasted in the form of heat. In this paper, we quantitatively and qualitatively analyze the current state of AHF emissions of the top three cloud service providers (i.e., Google, Azure and Amazon) according to their average energy consumption and the global distribution of their DCs. In this study, we found that Microsoft Azure DCs emit the highest amounts of AHF, followed by Amazon and Google, respectively. We also found that Europe is the most negatively affected by AHF of public DCs, due to its small area relative to other continents and the large number of cloud DCs within. Accordingly, we present mean estimations of continental AHF density per square meter. Following our results, we found that the top three clouds (with waste heat at a rate of 1,720.512 MW) contribute an average of more than 2.8% out of averaged continental AHF emissions. Using this percentage, we provide future trends estimations of AHF densities in the period [2020–2100]. In one of the presented scenarios, our estimations predict that by 2100, AHF of public clouds DCs will reach 0.01 Wm−2.

Methods #

• The paper focuses on the three hyperscale cloud providers and uses their average PUE to calculate the efficiency:
  • Google PUE = 1.10 = 9.1% energy waste.
  • Microsoft PUE = 1.25 = 20% energy waste.
  • Amazon PUE = 1.14 = 12.3% energy waste.
• The paper then calculates how many data centers each cloud provider has by equating a zone to a data center.
• It applies an average of 11.6 MW per data center to calculate a total waste heat for each data center for each cloud provider, based on those above efficiencies.

Notes #

The paper provides a good background about why this is a relevant topic to consider.

• Waste heat is the non-useful part of any heat generated by an industrial process.
• Anthroprogenic heat flux (AHF) is waste heat generated by human activity that is emitted into the environment and the atmosphere.
• AHF contributes to global warming.
• Computers generate heat as part of their normal operating processes. This is why data centers provide a cool environment to maintain the most efficient operating conditions for those servers.
• Data centers are not 100% efficient, so they generate waste heat, which causes AHF, and can therefore be linked to global warming.

It’s fairly obvious that heat is a byproduct of normal data center operations – indeed, that is why water is a major part of the environmental footprint of a data center (although mostly in thermal power generation rather than cooling). So I agree that this is a relevant topic to consider – how much waste heat is generated by a data center?

This paper hinges on two assumptions:

1. PUE measures energy, but does that include non-electrical energy? PUE is a ratio of how much energy is delivered to the computing equipment vs “everything else”, which includes the energy that goes into cooling, lighting and electrical distribution. For most data centers, that “energy in” is going to be electricity. It’s electricity that powers the servers, lights, and the cooling systems. Heat is a byproduct of the computer equipment processing, and electricity is used to cool the ambient environment (often by chilling water and pulling the air through the cooling system).

   It seems unlikely that a PUE of 1.25 means all of that .25 should be represented as heat. It depends on the cooling mechanism. For example, we know that some newer data centers make use of the ambient air to cool the facility through adiabatic cooling.

2. The paper assumes that every data center can be represented by an energy average of 11.6 MW. This figure is based on analysis from 2008. Although they validate this against Google’s annual report from 2019, this misses some complicating factors.

   First, zones are not necessarily complete data centers – in some cases they are, but they may be data halls within other facilities. This can be seen from the fact that Google does not disclose PUE figures for every region e.g. it has 3 zones in London, UK, but publishes no figures on its PUE dashboard. This is because Google has no owned facility in London – it uses colo providers instead.

   Second, this assumes that every DC is the same size. Some of the largest hyperscale facilities are huge warehouses much larger than 11.6MW. AWS states their preferred facility size is 25-32MW.

Conclusions #

Prior to reading this paper I had not thought about the environmental impact of waste heat and the resulting anthropogenic heat flux. It clearly has an environmental impact because we see this elsewhere (such as in the context of urbanization). However, I am not convinced that the assumptions made in this paper about cloud data centers are valid.

If the numbers suggested as correct, cloud data center make up only around 3% of the total global AHF footprint. The paper suggests that this could be offset by the negative waste heat from the use of renewable electricity, so that the impact nets to zero:

For example, if some CSPs are partly using renewable energy, this will reduce the greenhouse effect, since part of this energy is consumed by the DCs operations rather than contributing to greenhouse effects. That is, reducing the natural heat effect, while increasing the AHF, may lead to Zero net change. In optimistic scenarios, such climate-aware practices may indeed reduce the averaged total AHF values.

It’s difficult to know if these numbers are valid. On the one hand, they underestimate the size of many hyperscale data centers by 2-3 times. On the other hand, they don’t consider how cloud datacenters are architected in terms of owned facilities vs deployment within colo, which may be much smaller. Does that just average out to 11.6MW? There’s an additional question about the assumed efficiencies and whether all of the “excess” non-IT energy can be represented as heat.

I also ignored the projections made to 2100 because any projections (on any topic) made out more than 2-3 years lack credibility.

As such, my conclusion is that this paper (and AHF from cloud data centers) in inconclusive. If 3-6% of AHF is represented by cloud data centers I’m not sure if that is a problem, especially if renewables can offset it. But there is too much uncertainty in the calculations to be sure.

Either way, AHF is a new impact I’d not read about before, so it was an interesting read.