1 of 9

Timeshifting computation to reduce the carbon emission of software

Dr. Anne Hartebrodt

Biomedical Network Science Lab

Department Artificial Intelligence in Biomedical Engineering

Friedrich-Alexander-Universität Erlangen-Nürnberg

https://www.bionets.tf.fau.de

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

2 of 9

Computing emits carbon…

Greenhouse gas protocol:

Scope 1: On-site operations
Scope 2: Energy-consumption on-site
Scope 3: Supply chain/ Embedded emissions

Carbon intensity:

How much carbon is emitted per KWh
Fluctuates significantly throughout the day and geographically

-> There is carbon saving potential by shifting the computation to a suitable time or location

…. 2.8 - 3.9% of global greenhouse gas emissions [1].

(Foto: Erich Malter/RRZE)

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

3 of 9

How does that relate to cloud computing?

Timeshifting: Delay computation until CO_2eq/KwH falls below a certain threshold.

Why?:

CO2 intensity of electricity

production fluctuates throughout

the day.

Pros:

Can save up the 8% CO₂ emission [2]

Cons:

Assumes that you are not at

full capacity all the time.

Problems:

Prediction horizon (18:00)
Data availability
Data granularity

Time shifting

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

4 of 9

How does that relate to cloud computing?

Location shifting: Move the computation to a

geographic location where the energy is sources

from less carbon intensive energy sources.

Pros:

Could technically half the emission? [2]
No time delay.

Cons:

Assumes that you can move

your data (privacy, volume, …).

Problems:

Prediction horizon (18:00)
Data availability
Data granularity

Location shifting

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

5 of 9

<code></green> Software

API+ python client

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

6 of 9

<code></green> Software

API+ python client

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

7 of 9

<code></green> Software

Currently: ‘Green-ness’ measures as the percentage of wind and solar from total energy

other low carbon sources: hydro and nuclear.
Optimal: Use carbon-intensity.
Also possible: Use price (correlates heavily with renewables)

Improve parametrization:

Run time can be estimated for standard workflows based on previous runs.
Potential delay can be projected.
Looking for trace files, reports, etc to gain understanding for standard procedures

Total run time, CPU, GPU, Memory usage, workflow steps/tools.

Improve usability

Estimate realistic gains in CO₂ efficiency.

Work in progress

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

8 of 9

Acknowledgements

Biomedical Network Science Lab @AIBE @FAU

Prof. Dr. David B. Blumenthal

Students

Shubh V. Jain

Soham Ekbote

Hyunjung Wang

Pratiksha Manik Patil

Contact:

Anne Hartebrodt

anne.hartebrodt@fau.de

Send me:

Machine readable trace files, execution reports. Add in subject: [EGD-carbon-aware]
Questions!
Ideas!

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50

9 of 9

References

Image sources:

https://www.forbes.com/home-improvement/solar/american-solar-panel-manufacturers/

https://www.pv-magazine.com/2022/09/15/germany-to-tender-1-5-gw-of-additional-solar-power/

https://newtransparency.entsoe.eu

https://app.electricitymaps.com/map

https://smard.de

[1] The real climate and transformative impact of ICT: A critique of estimates, trends, and regulations. Charlotte Freitag, Mike Berners-Lee, Kelly Widdicks, Bran Knowles, Gordon S. Blair, and Adrian Friday, https://doi.org/10.1016/j.patter.2021.100340

[2] Dodge, J. et al. Measuring the Carbon Intensity of AI in Cloud Instances. in 2022 ACM Conference on Fairness, Accountability, and Transparency 1877–1894 (ACM, 2022). doi:10.1145/3531146.3533234.

08,70

06,00

05,00

00,00

07,70

08,90

15,50

00,30

15,50