Embodied carbon plugin

[jawache]

Several years ago Benjamin Davy from Teads created a dataset to model the emissions of servers. Prior to this work there wasn't really any simple way to estimate the energy/carbon profile of a server or CPU in a ubiquitous way other than asking for a manufacturer to disclose the data which they only did under NDA and usually with key customers.

The approach he took was to build a simple model which could then be extrapolated oiut and applied to any CPU/server. It made estimating the energy consumption of CPUs trivial, as long as you kew the TDP of a CPU you could estimate the energy consumption from CPU utilization.

In the impact framework world the process of taking an observation and turning it into an impact is called induction. So we induce cpu utilization into energy.

For energy these simple co-efficiants are what we call the teads curve.

It's a very rough estimate, but what it has for itself is broad application. You can apply this to everything and people do, it's used in the cloud-carbon footprint tool, which is used in climatiq which is used in dynatrace and I've seen it applied in about a dozen closed source internal tools at a variety of organizations. It might not be accurate on a granular level but when the next best thing is "no-data-at-all" it competes very well.

Almost every case study created for the SCI used the co-efficients created by Benjamin in this dataset.

To make calculating the SCI easier with the Impact Framework one of the first plugins we created was what we commonly refer to as the Teads Curve: https://github.com/Green-Software-Foundation/if-unofficial-plugins/blob/main/src/lib/teads-curve/README.md. This helps convert a cpu/utilization into an cpu/energy figure.

### Embodied Carbon Plugin

What I would like someone to propose for the hackathon is to take the same approach and create a plugin which outputs cpu/embodied-carbon.

Read the work from Benjamin here: https://medium.com/teads-engineering/building-an-aws-ec2-carbon-emissions-dataset-3f0fd76c98ac

The energy model is leverages in many places but there is another naive model to estimate embodied emissions from a server.

Assume a base total embodied carbon of 1000kg, then for each component listed below add that amount of carbon to the total.

The inputs to this plugin might be something along the lines of:

• server/dram
• server/cpu-count
• server/hdd-count
• server/sdd-count
• server/gpu-count

Note

I would enhance the existing cloud metdata data plugin to output those values for a given cloud instance type. It currently outputs cpu/thermal-design-power to support the teads-curve plugin, it can be enhanced to output the values above also (we maintain the tdp manually, so this would require some manual effort to capture the above)

Then this plugin would compute the embodied carbon of the sever according to those inputs and the model and output

• server/embodied-carbon

Which can then be used in downstream models like the sci model to compute carbon per user or some other metric.

Currently the best method we have for adding embodied carbon to a pipeline is the sci-m model, it's very clunkly and requires manually adding in embodied carbon values for every single server you want to measure. A plugin like the above would automate and simplify the whole process. If it was easier, more people would compute and report the embodied carbon of their compute.

Important

Benjamin moved on evolved the work with Boavizta who surface it via a free API. We've built a plugin to that API also which you can use for comparison/reference. https://github.com/Green-Software-Foundation/if-unofficial-plugins/blob/main/src/lib/boavizta/README.md

References

• Existing Teads Curve Plugin: https://github.com/Green-Software-Foundation/if-unofficial-plugins/blob/main/src/lib/teads-curve/README.md
• Existing Cloud Meta Data Plugin: https://github.com/Green-Software-Foundation/if-plugins/blob/main/src/lib/cloud-instance-metadata/README.md
• Original Work From Teads: https://medium.com/teads-engineering/building-an-aws-ec2-carbon-emissions-dataset-3f0fd76c98ac
• Boavizta IF Plugins: https://github.com/Green-Software-Foundation/if-unofficial-plugins/blob/main/src/lib/boavizta/README.md

https://github.com/Green-Software-Foundation/if-unofficial-plugins/blob/main/src/lib/teads-curve/README.md

https://medium.com/teads-engineering/building-an-aws-ec2-carbon-emissions-dataset-3f0fd76c98ac

[github-benjamin-davy]

Hello @jawache embodied emission data is scarce and what I suggested a while back was quite simplistic. I would encourage using the Boavizta work that is improving a lot what the initial dataset attempted naively :) https://doc.api.boavizta.org/Explanations/services/cloud/
Also, we have to keep in mind that the cloud infrastructure that we are evaluating when looking at our bills is only the visible part of the required hardware used to run the services, we would also need to include an estimation of the hardware used for the control plane and the surrounding equipment/building.

[da-ekchajzer]

We would be happy to collaborate on this specific issue. Do not hesitate if you have specific questions. We are trying to develop the Boavizta's international group. If you think it is relevant, we could collaborate with GSF through this channel.

Also, we have to keep in mind that the cloud infrastructure that we are evaluating when looking at our bills is only the visible part of the required hardware used to run the services, we would also need to include an estimation of the hardware used for the control plane and the surrounding equipment/building.

We are working on a new feature to include the missing parts of the cloud infra. We hope to publish this (along with all the methodology) in a peer review conference.

Concerning the power draw curves (what you call TEADS curves), they have been improved in the API. See : https://doc.api.boavizta.org/Explanations/services/cloud/#consumption-profile. As you already know, we are working on Energizta to improve them. Once version 1 is launched, the GSF will be a great help to push you community to take part in the collect of server consumption data.