VEEP: Video Encoding Energy and CO₂ Emission Prediction

In the context of rising environmental concerns, this paper introduces VEEP, an architecture designed to predict energy consumption and CO2 emissions in cloud-based video encoding. VEEP combines video analysis with machine learning (ML)-based energy prediction and real-time carbon intensity, enabling precise estimations of CPU energy usage and CO2 emissions during the encoding process. It is trained on the Video Complexity Dataset (VCD) and encoding results from various AWS EC2 instances. VEEP achieves high accuracy, indicated by an 𝑅2-score of 0.96, a mean absolute error (MAE) of 2.41 × 10−5, and a mean squared error (MSE) of 1.67 × 10−9. An important finding is the potential to reduce emissions by up to 375 times when comparing cloud instances and their locations. These results highlight the importance of considering environmental factors in cloud computing.

1. Armin Lachini, Manuel Hoi, Samira Afzal, Sandro Linder, Farzad Tashtarian, Radu Prodan, Christian Timmerer
Institute of Information Technology (ITEC), Alpen-Adria-Universität Austria
[email protected] | https://athena.itec.aau.at/
VEEP: Video Encoding Energy
and CO₂ Emission Prediction

2. 1
Motivation & Objective
66% of internet traffic are video content
Encoding is an energy intense task
Predict cloud-instance energy-consumption for encoding
Calculate CO₂ emissions for specific cloud-region

3. 2
Architecture – Overview

4. 3
Architecture – Video Analyzer

5. Video Analyzer
[1] VCA GitHub Repository: https://github.com/cd-athena/VCA
Video Complexity Analyzer (VCA)[1]
Open source tool
YUV or Y4M
Outputs complexity features
4

6. 5
Complexity Features
E = 137
L = 156

7. 6
Complexity Features
E = 137
L = 156
avgU = 149
avgV = 144

8. 7
Complexity FeaturesE = 137 L = 156 avgU = 149
energyU = 92avgV = 144
energyV = 44

9. Video Analyzer
[1] VCA GitHub Repository: https://github.com/cd-athena/VCA
Video Complexity Analyzer (VCA)[1]
Open source tool
YUV or Y4M
Outputs complexity features
8

10. 9
Architecture – Video Analyzer

11. Architecture – Energy Predictor
10

12. Energy Predictor
Trained on VE-Match[2] dataset
Encoding data of 500 videos on AWS EC2 instances
Data on encoding CPU energy usage
11
[2] https://dl.acm.org/doi/abs/10.1145/3593908.3593943

13. Energy Predictor
Trained on VE-Match[2] dataset
Encoding data of 500 videos on AWS EC2 instances
Data on encoding CPU energy usage
12
[2] https://dl.acm.org/doi/abs/10.1145/3593908.3593943

14. Architecture – CO₂ Data Source
13

15. Architecture – CO₂ Data Source
14

16. CO₂ Data Source
Electricity Maps[3]
Live data for carbon intensity
Accounts for imports/exports
15
[3] https://www.electricitymaps.com/

17. Annual Overview
CO₂ emissions over the year 2023 in Austria.
16

18. Daily Overview
17

19. Architecture – CO₂ Calculator
18

20. Architecture – CO₂ Calculator
19

21. Architecture
20

22. Analysis
21

23. Methodology
Instance Types
Types optimized for different usecases
Different hardware configurations
22

24. Methodology
Instance Types
Types optimized for different usecases
Different hardware configurations
Codec
AVC/H.264
Regions
Global selection of countries
Diverse mix of energy mixes
23

25. Emissions per Instance Type
CO₂ emissions for encoding Wood_s001 on different
instances in India
24

26. Emissions per Instance Type
CO₂ emissions for encoding Wood_s001 on different
instances in India
25
11x

27. Emissions per Country
CO₂ emissions for encoding Wood_s001 on c5.large
in different countries
26

28. Emissions per Country
CO₂ emissions for encoding Wood_s001 on c5.large
in different countries
27
37x

29. Conclusion
Core goal
Predict CPU energy
Calculate CO₂ emissions
Achievements
Energy-prediction with R²-score of 0.96
CO₂ emission reduction up to 405 times
Future work
Consider findings during scheduling of CPU intense
jobs
28

30. Thank you
Have a
great day
ahead!
Paper Link
https://doi.org/10.1145/3652104.3652528
Alpen-Adria-Universität Austria
Institute of Information Technology (ITEC)
Armin Lachini
[email protected]