Home|All Posts| Blog

Carbon Accountability for Cloud Services and the need for it in 2021

Nour Rteil | July 13, 2021

Carbon Accountability for Cloud Services and the need for it in 2021

As our demand for online services grows rapidly, data centres must make vital changes to meet this demand sustainably and reduce the footprint of their cloud services sooner than later. According to a recent report by International Data Center (IDC), cloud computing can potentially eliminate 1 billion metric ton of CO2 emission from 2021 to 2024. Considerable work is being done already and many data centres are in the process of incorporating zero-emission goals in their business strategies.

While this is great news, users now need more than just promises that are set 5,10 and 20 years from now. An imperative step in reducing the climate footprint, that can be achieved today, is carbon accountability and transparency. Users should be able to evaluate the carbon footprint of their services hosted on different platforms and choose the service provider that best fits their business needs and sustainability goals.

Cloud providers, however, are secretive when it comes to their energy use and greenhouse gas emissions according to the latest report by Greenpeace, but this may be changing with multiple initiatives having been set out to evaluate the carbon cost of the cloud. In 2020, Microsoft launched a sustainability calculator allowing Azure cloud customers to gain insight into their carbon emissions across all scopes (including scope 3).

The methodology behind this calculator is explained in their recent white paper, where the carbon emissions are quantified based on aggregated IT hardware emissions per region and allocated to customers based on their usage per service per region. Though this is a great initiative by Microsoft, using monthly usage as a proxy could be tricky since usage is calculated for virtual resources, without considering load and the energy consumption impact of physical servers. A virtualised server could be running at 70% but its physical server could be utilised only 15%. Similarly, a virtualised server could be running at 20% while its physical server is utilised 50%. As a result, 2 instances running at different loads could have the, inaccurately, the same carbon footprint.

Another tool that allows users to estimate the carbon footprint of cloud services is the open-source Cloud Computing Footprint. This tool gives estimates to services hosted on AWS, GCP, Azure, and has alternative data approaches for other providers. It is based on Etsy’s Cloud Jewel methodology developed in 2020. The methodology allocates the following point estimates, with no confidence levels, as to not provide false precision: 2.1Wh per virtualised CPU, 0.89 Wh/tbh for HDD storage, and 1.52 Wh/tbh for SSD storage. The energy per virtualised CPU is estimated based on the average of the power values that are available in SPEC power database, filtered to servers that they deemed likely to be similar to the hosting data centre’s servers, and an average utilisation rate of 45%. Storage point estimates are based on the industry-wide estimates from the US Data Usage Report. Networking and RAM are not accounted for in this methodology due to a lack of information on the energy impact of these components.

Cloud services carbon estimates are then calculated according to this equation:

(Cloud provider service usage) x (Cloud energy conversion factors [kWh]) x (Cloud provider Power Usage Effectiveness (PUE)) x (grid emissions factors [metric tons CO2e])

Whilst the methodology used here is robust, the utilisation rate and power consumption of the physical server should be considered instead of the virtualised instance to avoid wrong assumptions, as explained above. In addition, the energy estimate of 2.1Wh for a virtualised CPU is debatable, especially that each server, and even each hardware configuration for the same server, has a different energy draw depending on its model and specs.

The variation of energy consumption from one generation to the other, and one configuration to the other, varies significantly as demonstrated in this IEEE paper. Averaging out the power consumption of different servers can yield huge error margins (this is where a tool like Interact would be highly needed to accurately estimate each server’s power consumption). To add to that, Google and other hyperscalers like Facebook and Microsoft use their own custom servers that have different power consumptions than vendor-based servers listed in the SPECpower database, for which we don’t have power information.

Surely, carbon footprint calculators like these help users acquire knowledge of a data centre’s carbon footprint and allow them to take the proper actions thereafter. Despite some imprecise estimations, these tools will encourage others to do the same and take account of their carbon footprint across all scopes (1,2 and 3).

There is, however, a big need for more transparency on life cycle analyses especially by the big players (that host 60% or more of the global cloud services). This can benefit the community and enable the creation of a unified benchmark to establish footprint estimates with high confidence rates so users can take educated decisions when it comes to choosing the right service provider and hosting location.