fdiblen Environmentally sustainable research computing
Research computing has a real carbon footprint - large simulations, ML training runs and always-on services consume energy at a scale funders and institutions increasingly ask about. The GREENER principles for computational science frame the practice: Governance, Responsibility, Estimation, Energy and embodied impacts, New collaborations, Education and Research. For an agent the working core is: measure first, reduce second, schedule smart, report honestly.
Measure before optimizing
Never guess at footprint - estimate it:
- CodeCarbon instruments Python code with a few lines (or a CLI wrapper) and estimates energy plus location-adjusted CO2e; add it to representative runs, not every run.
- For non-Python or cluster workloads, estimate from job accounting (CPU/GPU hours x hardware power draw x facility PUE x grid carbon intensity) and state the assumptions.
- Record estimates alongside results the same way runtimes are recorded, so the cost of a paper's computations is reportable.
The Software Carbon Intensity (SCI) specification (an ISO standard) gives a defensible formula when a formal number is needed: operational plus embodied emissions per functional unit.
Reduce
Order interventions by leverage, and quantify the win when possible:
- Compute less: cache intermediate results, avoid re-running unchanged pipeline stages (rseng-workflows), kill zombie jobs, right-size parameter sweeps before launching them.
- Compute efficiently: profile first (rseng-performance-profiling) - a 5x speedup is usually a ~5x energy cut; use appropriate precision; prefer vectorized/compiled paths in hot loops (rseng-language-guides).
- Match hardware to the job: GPUs are more energy-efficient than CPUs for the workloads that suit them (rseng-gpu-computing) and wasteful for the ones that do not; do not reserve more nodes, memory or walltime than the job uses.
- Store less: data has a footprint too - prune intermediates, compress archives, apply retention rules (rseng-data-management).
Schedule smart
Grid carbon intensity varies by hours and by region. Carbon-aware scheduling shifts flexible batch work to cleaner windows:
- CATS (Climate-Aware Task Scheduler) picks the lowest-carbon start time for a job of a given duration on UK-grid data; the same delay-tolerant principle applies anywhere batch work is flexible.
- Cloud users can choose lower-carbon regions for flexible workloads; cluster users can prefer off-peak windows where the operator exposes them.
Report and advocate
- Include a brief compute-footprint statement in papers and READMEs for compute-heavy projects (estimated kWh/CO2e and the estimation method) - normalize the practice.
- When proposing CI pipelines, keep them lean: cache dependencies, skip redundant matrix entries, avoid scheduled jobs nobody reads (rseng-ci-cd).
- Educate while doing: a measured number ("this sweep emitted an estimated 12 kg CO2e") lands better than generic advice.
Working with this skill
This skill is source-independent: its authority is the GREENER principles, the SCI specification and the tool documentation linked below.
Learn more (verified):
- https://www.nature.com/articles/s43588-023-00461-y - GREENER principles for environmentally sustainable computational science
- https://codecarbon.io - CodeCarbon energy/CO2e estimation
- https://greensoftware.foundation/standards/sci/ - Software Carbon Intensity specification
- https://github.com/GreenScheduler/cats - Climate-Aware Task Scheduler
Related skills
Check whether any of these applies before moving on:
- rseng-ci-cd - lean pipelines waste less compute
- rseng-data-management - storage retention has a footprint
- rseng-gpu-computing - matching hardware to workload efficiency
- rseng-hpc-computing - right-sized resource requests save energy
- rseng-performance-profiling - speedups cut energy roughly proportionally
- rseng-workflows - caching avoids recomputing pipeline stages