1 of 18

On the Limitations of Carbon-Aware Temporal and Spatial Workload Shifting in the Cloud

0

Thanathorn Sukprasert, Abel Souza, *Noman Bashir, David Irwin, Prashant Shenoy

University of Massachusetts, *Massachusetts Institute of Technology

Eurosys’24 April 25

2 of 18

Introduction

Computing demand is increasing at a rapidly accelerating rate

1

training compute has grown by a factor of 10 billion since 2010

[1]

[1]: https://airtable.com/appDFXXgaG1xLtXGL/shrBucz1oynb4AUab/tblhmFk3gP7psWh3C

3 of 18

Introduction

Datacenters’ energy consumption is also rapidly accelerating

2

[2] iea.org

[2]

Energy Consumption (TWh)

4 of 18

Introduction

Datacenters’ carbon footprint is also rapidly accelerating

3

[3]: journal.uptimeinstitute.com

[4]:ICT Sector Electricity Consumption and Greenhouse Gas Emissions – 2020 Outcome

[4]

[3]

Average Annual PUE

While the computing demand continues to grow at a rapidly accelerating rate, the power usage effectiveness, PUE, however, is stalling. This causes the datacenters’ carbon footprint to also be increasing at a rapidly accelerating rate. This figure shows carbon emissions from ICT sector from 2007 to 2020. It can be seen that since 2015, the total ICT emissions have increased by around 5%.

As the computing demand grows at a rapidly accelerating rate,

the datacenters’ carbon footprint is also rapidly accelerating. This figure shows that the power usage effectiveness, PUE, is stalling and the growth in computing demand outgrows the gains in energy-efficiency

This figure shows that the power usage effectiveness, PUE, is stalling, however, as we discussed, the computing demand is increasing at accelerating rate. Since the growth in computing demand outgrows the gains in energy-efficiency, there is a rise in carbon footprint in the datacenters.

This figure shows carbon emissions from ICT sector from 2005 to 2020. It can be seen that since 2015, the total ICT emissions have increased by around 5%.

5 of 18

Introduction

4

Computing Demand

Energy Demand

Carbon

Emissions

Need to make

cloud computing

sustainable

Optimize for carbon efficiency

6 of 18

Carbon-Efficiency

5

PUE ~1: highly energy efficient

PUE ~2: highly energy inefficient

Coal powered

Solar powered

Highly carbon inefficient

Highly carbon efficient

Highly energy-efficient datacenters can still be carbon-inefficient

7 of 18

Optimizing for Carbon-Efficiency

6

California

Region R

Sweden

Harness the variations of carbon intensity across time and locations

Temporal Shifting

Spatial Shifting

Next, I am going to explain how we can optimize for carbon efficiency. Let’s look at the carbon intensity in California where there is a low-carbon period during the day for its solar availability and high carbon period at night when there is no sun. Now let’s suppose we have a workload that needs to be executed. Instead of running the workload right away, we can schedule the workload during the low-carbon periods, this is called temporal shifting.

Furthermore, let’s supposed we have two regions, region R with high carbon intensity for the electricity is coaled powered and another region, Sweden, with low carbon intensity for the electricity is hydroelectricity. Instead of running the workload at the high carbon region, we can migrate the workload to the low-intensity region, this is called spatial shifting.

This can be seen that optimizing for carbon efficiency can be done by harnessing the the variations of carbon intensity across time and locations.

8 of 18

Research Question

Variations in the carbon intensity worldwide, both temporally and spatially
The potential benefits of spatiotemporal shifting are unclear

7

9 of 18

Our Work

Quantify the potential benefits and limitations of carbon reduction from temporal and spatial workload optimization

123 regions from Electricity Maps
Cover 99 datacenters
From 2020-2022

8

10 of 18

Spatial Workload Shifting

Migrate a 1-hour batch job the lowest carbon intensity location globally: Sweden

9

96% carbon reduction!

🌎

Let’s start with spatial workload shifting by migrating a 1-hour batch job the lowest carbon intensity location globally, Sweden. The average carbon intensity across all 123 regions is 368 gCo2Eq… and by migrating all the job to Sweden, we can reduce the carbon emissions of the job to about 16 gCOeq.

What this means is that if everyone migrate their workload to sweden, we can reduce the carbon emissions of the whole world by 96%.

The figure on the right shows an average emission of a 1-hour batch job for each geographical region before getting migrated to Sweden.

This result shows that there is a seemingly a large potential for carbon reduction for spatial shifting

average carbon intensity globally is 368", mention “the average carbon intensity globally, across all the 123 regions, is 368 gCo2Eq…

11 of 18

Spatial Workload Shifting

10

Migrating once yields most of the carbon reduction

12 of 18

Spatial: Capacity Constraints

However, every region has a capacity constraint

11

Capacity constraints dimmish the carbon reduction opportunity

13 of 18

Temporal: Deferrability

Deferrability: Delay the job start time

12

CO₂Intensity

time

Deferrability is beneficial for short jobs

24H Slack

Slack

14 of 18

Temporal: Interruptibility

Interruptibility: suspend and resume

13

24H Slack

CO₂Intensity

Interruptibility augments the carbon reductions for long jobs

time

15 of 18

Temporal Workload Shifting

High variance 🡪 high reduction
But high variance regions are generally low-carbon regions

14

Low variance regions have high emissions but cannot enjoy the temporal shifting benefits

16 of 18

Increasing Renewables

15

Temporal: California

Spatial: California

Carbon-agnostic scheduling also results in low carbon emissions

17 of 18

Conclusion

The practical carbon reductions are far from ideal
Sophisticated migration techniques yields small additional carbon reduction
Temporal shifting has limited gains for long jobs and low-variance regions
The benefits of carbon-aware workload scheduling will decrease as the world's energy supply becomes “greener”

16

18 of 18

Thank you�Questions?

17

Thanathorn Sukprasert

tsukprasert@umass.edu

On the Limitations of Carbon-Aware Temporal and Spatial Workload Shifting in the Cloud