5a. Design for efficient cloud architecture

Hosting services on ‘cloud’ can bring efficiencies relative to on-premises infrastructure. However, there are still major environmental impacts that can come from the electricity consumption of servers, energy and/or water used for server cooling and the embodied impacts coming from the manufacture, distribution and disposal of buildings, cooling equipment and servers.

Lifecycle phases

Alpha Beta Live

Actions

(i) Design the best fit cloud services

This might mean managed services, which can achieve high utilisation and efficiencies through shared resources and simplify the task of updating software.

(ii) Run workloads in places and at times when low-carbon electricity is available

Cloud data centres are located in many countries with highly varied electricity grids. If available, move workloads to a hosting provider that uses electricity from low carbon sources like wind, solar, geothermal, hydro, or nuclear.

In addition, if possible, design for ‘carbon-aware’, and run workloads that are not time-critical at times when there is low-carbon electricity available at a given location.

More detailed guidance on cloud architecture & sustainability

There is guidance relating specifically to the sustainable design of cloud architecture available from the following sources:

'Well-Architected Frameworks' from major public cloud providers

Public cloud providers Amazon, Microsoft and Google all have their own ‘Well-architected frameworks’ (WAF) which cover cover guidelines for general best practice when it comes to building and maintaining cloud architecture. There are many commonalities as well as differences between the respective frameworks.

Measurement

Cloud hosting providers may provide reporting dashboards, but caution is required as in many cases there may be a substantial difference between reported emissions and actual consumption at data centres.

There are a range of questions that we can ask when we view reporting cloud providers:

  • Are energy usage figures provided or are spend-based proxies used to estimate carbon emissions?

  • For Scope 2, are location-based electricity figures provided or are market-based emissions provided (that may rely on the purchase of green electricity through ‘Power Purchasing Agreements (PPAs))?

  • Do figures follow the GHG protocol and include Scope 3 emissions alongside Scopes 1 and 2? (The embodied emissions may include up to half the carbon caused by cloud computing)

  • Is reporting using offsets to claim that there is zero or extremely low carbon associated with cloud activities?

Using a tool to estimate cloud carbon footprint

Cloud Carbon Footprint is an open-source tool to help decipher public cloud. It helps look at billing statements from a provider line by line (the log of activities, if tags are assigned to services to provide meaningful data that can interpret). For storage, compute, hardware etc it estimates how much electricity use and also carbon based around location.

If you are able to get energy usage data, you can use services like Electricity Maps to then calculate the emissions emitted in relevant region of the world.

Further Reading

Creating and Implementing a Cloud Hosting Strategy, GOV.UK

Cloud Guide for the Public Sector, GOV.UK

Five Things to Know About Cloud Computing GHG Emissions, Carbometrix

Building Green Software, Currie, Anne; Hsu, Sarah; Bergman, Sara

5b. Design for optimised on-premise systems

The goal of this guideline is to set up any on-premises systems that your service uses to be as efficient as possible.

[Detail to be added to this guideline]

Lifecycle phases

Alpha Public Beta

Actions

(i) Run systems at higher utilization to improve efficiency

(ii) Use well optimised open source libraries or products

5c. Be smart with AI

AI may offer us the opportunity to make businesses more resource-efficient, reducing costs, emissions, and waste through monitoring of energy consumption, hardware utilization and data storage. It may also bring environmental benefit through other specific applications that bring environmental benefit, such as analysing data sets to better support services (e.g. satellite or weather data). But AI is also resource-hungry and training models consume large amounts of water and produces a carbon footprint.

Actions

(i) Choose the appropriate size model

Choose the appropriate AI model for the service's needs. Small AI models can be trained for a specific task, are cost effective, quick (as there is less data for them to reference) and they can be run on basic hardware (such as a PC or phone.

(ii) Reuse existing models where possible

Consider building on an existing model rather than starting from scratch, and use AI to analyse code performance and then tidy to lower environmental impact.

(iii) Carefully choose supplier and region

Careful selection of the right supplier and location is important. Training models in low-carbon regions or times of day can reduce the carbon footprint. Don't forget about the other environmental impacts of AI, especially water use (but even this can be complex - water use in Scotland will have a different impact to Suffolk, for example).