Teads AWS Estimation Model
Goal: Estimate the carbon footprint of using AWS EC2 resources.
AWS EC2 Carbon Footprint Dataset: Google Sheets spreadsheet
Methodology
Tool turbostress was used to perform several stress tests to simulate different workloads and to report power consumption measures using RAPL.
Notes:
- Available Intel-based bare-metal instances on AWS were tested: c5, m5, r5, m5zn, z1d, i3, c5n
- Other platforms (AMD, ARM) weren't measured, but estimated based on known TDP values.
- PUE is included in the carbon estimates (AWS: 1.2).
Assumptions:
- billing reports reflect the actual usage of resources
- a bare-metal instance is cut into virtual instances linearly and all instances are used by other tenants (consumption of idle is split to all)
- generalization to all platforms can be made by using the results of the machine with the closest hardware and to estimate the consumption based on the TDP
Limitations:
- only CPU and RAM of EC2 instances are measured, other components like storage and network cards are missing
- consumption of resources that are used by AWS on top of EC2 like monitoring and logging is not included
- EC2 instance types use a processor of the same family, but sometimes different models, e.g. T2 (Xeon E5-2676 V3 & Xeon E5-2686 V4)
- Some EC2 instance types have burstable performance (e.g. using Intel Turbo Boost) → leads to inconsistencies
- CPU frequency is not always the same due to DVFS
- Platforms that are not Intel-based were not measure → Available TDP information was used to guess the energy consumption
🔗 References
Building an AWS EC2 Carbon Emissions Dataset | by Benjamin DAVY | Teads Engineering | Medium
Apidays LIVE Paris 2021 - Building an AWS EC2 Carbon Emissions Dataset By Benjamin Davy - YouTube