A Systematic Evaluation of the Potential of Carbon-Aware Execution for Scientific Workflows

This paper systematically evaluates the potential of carbon-aware execution strategies for scientific workflows, demonstrating that leveraging their inherent flexibility through temporal shifting and dynamic resource scaling can reduce carbon emissions by over 80% and 67%, respectively.

The Gist

Here is an explanation of the paper using simple language, analogies, and metaphors.

The Big Idea: Cooking Dinner When the Sun is Shining

Imagine you are a chef (a scientist) who needs to cook a massive, complex banquet (a scientific workflow) for a large group. This cooking process takes a long time, uses a lot of electricity, and produces a lot of smoke (carbon emissions).

Usually, chefs just start cooking whenever they feel like it. But what if the electricity grid is like a weather system? Sometimes the power comes from clean, free wind and sun (low carbon). Other times, it comes from dirty, smoky coal (high carbon).

This paper asks: What if we could wait to cook our dinner until the "clean energy weather" is perfect?

The researchers found that scientific workflows are actually perfect candidates for this. They are flexible, can be paused, and can be sped up or slowed down. By treating these computer tasks like a flexible dinner party, we can drastically cut down the pollution they create.

---

The Three Superpowers of Scientific Workflows

The paper highlights three special traits that make scientific workflows great at "green cooking":

1. Delay Tolerance (The "No Rush" Rule):

   • The Analogy: Unlike a fire alarm that must be answered instantly, a scientist analyzing DNA or galaxy images doesn't usually have a strict deadline. They can say, "I'll start this analysis tomorrow morning when the sun is shining," rather than "I must start it right now."
   • The Benefit: This allows us to Time Shift. We can pause the work and wait for the grid to be clean.

2. Interruptibility (The "Pause Button"):

   • The Analogy: Imagine you are baking a cake. If the power goes out, you don't throw the batter away. You put the bowl in the fridge, wait for the power to come back, and then finish baking. Scientific workflows work the same way. They save their progress to a hard drive, stop, and then resume later.
   • The Benefit: We can chop the work into small chunks. We run a chunk when the energy is clean, pause when it gets dirty, and resume when it's clean again.

3. Scalability (The "Team Size" Flexibility):

   • The Analogy: If you have a huge pile of dishes to wash, you can hire one person to do it slowly, or a whole team of 10 people to do it quickly.
   • The Benefit: When the energy is super clean (like a sunny afternoon), we can hire a huge team (use more computers) to blast through the work. When the energy is dirty, we hire fewer people and work slower.

---

The Experiments: What Happened When They Tried This?

The researchers took seven real-world scientific recipes (workflows) used in biology, astronomy, and earth observation and tested them in seven different countries (like the UK, USA, Germany, and Australia).

1. The "Wait and See" Test (Time Shifting)

They asked: If we just delay starting the whole job until the cleanest time of day, how much pollution do we save?

• The Result: In places with lots of wind and solar power (like California or the UK), they could cut emissions by over 80%.
• The Catch: In places that rely heavily on coal (like parts of South Africa), waiting didn't help much because the air was always "smoky."

2. The "Pause and Resume" Test (Interrupted Shifting)

They asked: What if we don't just wait to start, but we pause and restart the job multiple times throughout the day to catch every little burst of clean energy?

• The Result: This was even better! In California, they could cut emissions by 30–70% in just a single day.
• The Metaphor: It's like a surfer waiting for the perfect wave. Instead of waiting for one giant wave all day, they catch many small waves. This is much more efficient than just waiting for one big wave.

3. The "Change the Gear" Test (Resource Scaling)

They asked: What if we change the computer settings or the number of computers we use based on the energy?

• The Result:
  • Changing Gear: Computers have a "Performance" mode (fast but hungry) and a "Power Save" mode (slow but efficient). Sometimes, running slower actually saves more carbon because it aligns better with the clean energy available at that moment.
  • Changing Team Size: If the energy is super clean, they used more computers to finish the job faster. This "burst" of activity used the clean energy before it disappeared.
  • The Win: By smartly choosing how fast to run and how many computers to use, they reduced emissions by 67%.

---

The "Hidden Cost" of Pausing

The researchers were worried about one thing: Does pausing the computer to wait for clean energy waste energy?

• The Fear: If you pause a computer, does it still use electricity? Do we need to keep the hard drive spinning?
• The Reality: They did the math. Even if you pause a workflow for a whole day, the extra energy used to keep the data safe on the disk is tiny. It's like the difference between leaving a light on for an hour versus driving a car for a year. The savings from waiting for clean energy are massive compared to the tiny cost of waiting.

The Bottom Line

This paper proves that scientific computing doesn't have to be a dirty, polluting activity. By treating computer tasks like a flexible schedule rather than a rigid machine, scientists can:

1. Wait for the sun and wind to do the heavy lifting.
2. Pause their work when the grid gets dirty.
3. Speed up when the grid is super clean.

The Takeaway: We don't need to invent new, magical technology to save the planet. We just need to be a little more patient and flexible with our computers. By doing so, we can cut the carbon footprint of scientific research by 80% or more, turning the "smoky kitchen" of science into a "green garden."

Technical Summary

Here is a detailed technical summary of the paper "A Systematic Evaluation of the Potential of Carbon-Aware Execution for Scientific Workflows."

1. Problem Statement

Scientific workflows (e.g., in bioinformatics, astronomy, and earth observation) are computationally intensive, often running for hours to weeks on distributed clusters. This results in significant energy consumption and operational carbon emissions. While carbon-aware computing (aligning workloads with low-carbon energy availability) has been explored in general cloud contexts, its application to scientific workflows remains under-researched.

Existing sustainability efforts often focus on energy efficiency (e.g., DVFS, efficient scheduling), which faces hardware limitations and diminishing returns. The authors argue that scientific workflows possess unique properties—delay tolerance, interruptibility, scalability, and heterogeneity—that make them uniquely suited for carbon-aware execution (temporal shifting and resource scaling) to minimize emissions without necessarily sacrificing performance.

2. Methodology

The study employs a systematic, simulation-based evaluation using real-world data to quantify potential carbon savings.

• Workflows: Seven real-world Nextflow workflows were selected from diverse domains:
  • Bioinformatics: Chip-Seq, MAG, Nano-Seq, RNA-Seq, Sarek.
  • Astronomy: Montage.
  • Earth Observation: Rangeland.
• Infrastructure & Data:
  • Resources: Executions were traced on diverse hardware, including Google Cloud Platform (GCP) instances and various on-premise clusters (e.g., Olympus, Camelot, Atlantis) and edge servers.
  • Carbon Intensity (CI) Data: The study utilized high-resolution Average CI (grid-wide emissions) and Marginal CI (emissions of the specific source meeting additional load) data from Electricity Maps and WattTime.
  • Regions: Seven regions were analyzed: Great Britain, Germany, California, Texas, South Africa, Tokyo, and New South Wales.
• Carbon Estimation:
  • Operational Emissions: Calculated using a linear power model (via the tool Ichnos) applied to workflow traces, multiplied by CI data.
  • Embodied Emissions: Estimated using the Boavizta API, attributing hardware manufacturing emissions based on the runtime share of the device's expected lifespan (4 years for CPUs).
• Experimental Assumptions: To establish an upper bound on potential savings, the study assumed:
  1. Perfect knowledge of task runtimes and dependencies.
  2. Error-free CI forecasts.
  3. Infinite resource availability (no capacity constraints).
     Sensitivity analyses were later conducted to test the impact of forecast errors (5–15% for CI, 10–25% for runtimes).

3. Key Contributions

The paper makes three primary contributions:

1. Quantification of the Problem: It provides the first detailed carbon footprint estimates for seven popular scientific workflows across diverse infrastructures and regions, establishing baselines for operational and embodied emissions.
2. Systematic Evaluation of Strategies: It rigorously evaluates two carbon-aware strategies specifically for scientific workflows:
   • Temporal Shifting: Delaying execution (entire workflow) or pausing/resuming (interrupted workflow) to align with low-CI periods.
   • Resource Scaling: Dynamically selecting hardware nodes and adjusting processor governors (Performance vs. PowerSave) based on CI signals.
3. Open-Source Reproducibility: The authors released their simulation and analysis code to enable future research.

4. Key Results

A. Carbon-Aware Temporal Shifting

• Entire Workflow Shifting: Delaying the start of a workflow by up to 96 hours yielded significant reductions in regions with high renewable penetration.
  • Great Britain: Up to 80% reduction in emissions using Average CI.
  • South Africa: Minimal reduction (<15%) due to a coal-heavy, stable grid with low CI variability.
  • California & Texas: Significant reductions (up to 90% in specific months) when using Marginal CI, which captures periods of energy curtailment (near-zero CI).
• Interrupted Workflow Shifting: Pausing and resuming workflows to exploit multiple short low-carbon windows proved more effective than shifting the entire workflow.
  • In California, interrupted shifting achieved 30–70% reductions within a 6–24 hour window, outperforming entire workflow shifting in the same timeframe.
  • Overhead: The energy cost of storing intermediate data on disk during pauses was found to be negligible (increasing total energy by <6% even for data-intensive workflows).
• Sensitivity: The results remained robust even with forecast errors (5–15% for CI, 10–25% for runtimes), showing only minor deviations from optimal savings.

B. Carbon-Aware Resource Scaling

• Node Selection: Choosing specific hardware nodes based on their efficiency and local CI significantly altered emissions. For example, running tasks on older, more efficient nodes (like Camelot) sometimes yielded lower emissions than newer, more powerful nodes (Elysium) despite longer runtimes.
• Frequency Scaling (Governors):
  • Performance Governor: Faster runtimes but higher instantaneous power draw.
  • PowerSave Governor: Slower runtimes but lower power draw.
  • Result: In regions like Texas and California, using the PowerSave governor allowed workflows to run longer but align with low-carbon windows, reducing emissions by up to 67% compared to the Performance setting.
• Cluster Scaling: Increasing the number of nodes reduced runtime but slightly increased total energy consumption. However, by scaling up specifically during low-CI windows, overall emissions could be minimized.

C. Embodied vs. Operational Carbon

• The study found that while interrupting workflows increases CPU runtime (and thus the share of embodied carbon attributed to the task), the operational carbon savings vastly outweighed the increase in embodied emissions.
• For example, a workflow paused for 11 hours saw a negligible increase in embodied carbon (<4%) compared to the massive reduction in operational emissions.

5. Significance and Conclusion

This paper demonstrates that scientific workflows are ideal candidates for carbon-aware computing due to their inherent flexibility. The study proves that:

• Substantial Savings are Possible: Operational carbon emissions can be reduced by over 80% through temporal shifting and 67% through resource scaling.
• Signal Choice Matters: Using Marginal CI is critical for capturing curtailment events and achieving near-zero operational emissions in regions with high renewable variability.
• Interruptibility is Key: Allowing workflows to pause and resume offers greater flexibility and faster savings than simply delaying the start time.
• Trade-offs are Manageable: The overheads of storage and slight increases in embodied carbon are minimal compared to the operational benefits.

The authors conclude that by leveraging delay tolerance, interruptibility, and scalability, scientific communities can significantly reduce their carbon footprint without compromising scientific output, provided that scheduling systems are updated to utilize real-time carbon intensity data. Future work will focus on integrating these strategies into real-world schedulers with imperfect forecasts and resource constraints.