PaperlayerView exact revision on arXiv

Source-linked AI summary

GreenDirector: carbon- and water-aware workload placement for sustainable computing

Jaime Iglesias Blanco, Ignacio Heredia, María Castrillo, Andrei Tsaregorodtsev, Mazen Ezzeddine, Álvaro López García

arXiv:2609.12602v2cs.DC

TL;DR

Carbon-only accounting misses the local and seasonal water impacts of electricity, and prior operational placement approaches do not consistently unify carbon and standardized scarcity weighting. The paper introduces ES and GS for federated workload placement; evaluations report greener placement and quantified DIRAC carbon outcomes, alongside an explicit carbon-water trade-off.

  • Problem

    Operational workload placement lacks a practical standardized metric that unifies carbon and water impacts, although low-carbon electricity can be water-intensive and water stress is locally and seasonally dependent.

    Source: §1 · p. 1, §2.3 · p. 3

  • Method

    The paper combines cross-border electricity-flow tracing with monthly AWARE2.0 water-scarcity factors in ES, then uses hardware efficiency, PUE, and proportional environmental burden in GS for scheduling.

    Source: §3.1 · p. 3, §3.2 · p. 7, §6 · p. 19

  • Results

    DIRAC’s GS-ranked placement reduced carbon emissions by 42.8% in simulation and improved carbon efficiency by 49.5% in preliminary production, with higher water-scarcity impact when the lowest-carbon site had greater local water stress.

    Source: §6 · p. 19

  • Takeaways & Limitations

    The evaluations show greener providers can be prioritized in AI4EOSC without degrading user experience or scheduling latency, while DIRAC makes carbon-water trade-offs explicit for operator weight tuning.

    Source: §6 · p. 19

    The evaluated ES combines carbon and water only, covers operational electricity impacts rather than on-site cooling or embodied impacts, and depends on its input data’s resolution and assumptions.

    Source: §5.4 · p. 18

Abstract

from arXiv · show

The rapid growth of data center electricity demand, accelerated by AI, makes carbon-only accounting an incomplete measure of computing's environmental impact: low-carbon electricity mixes are often water-intensive, and the resulting harm depends on local, seasonal scarcity rather than on the volume of water consumed. We propose the Environmental Score (ES), a unified, dimensionless index in $[0, 100]$ that jointly captures the carbon footprint and the spatial-temporal, scarcity-weighted water impact of the electricity a workload consumes. It combines real-time, cross-border electricity flow tracing with monthly AWARE2.0 water-scarcity characterization factors, weighting global greenhouse-gas emissions together with the local, seasonal severity of water stress. Building on it, we define the Green Score (GS), a scheduling metric proportional to the useful computational work delivered per unit of real environmental impact, which also accounts for data center power and hardware efficiency. We add both metrics as a green-affinity feature to the GreenDirector schedulers of two production federated infrastructures, the AI4EOSC scientific cloud and the DIRAC workload management system. In AI4EOSC, a cluster-filling experiment over four pan-European providers shows that greener sites are filled first without degrading scheduling latency or end-user experience. In DIRAC, a trace-driven simulation of 133,631 jobs and a preliminary production deployment for the KM3NeT community reduce carbon emissions and improve carbon efficiency by about 40\%, while making the carbon-water trade-off explicit when the lowest-carbon site also carries higher water stress. The results show that hydrological stress can be dynamically weighted into workload placement in live, multi-tenant systems

1. Introduction

Carbon-only accounting misses the water impacts of electricity, which vary with local and seasonal scarcity and can conflict with low-carbon energy choices. The paper introduces environmental and scheduling scores and evaluates them in two production federated infrastructures.

1. Introduction

  • The Environmental Score (ES) combines carbon footprint with spatial-temporal, scarcity-weighted water impact from consumed electricity in a dimensionless [0, 100] index.It uses real-time cross-border electricity tracing and monthly AWARE2.0 factors.Source: §1 · p. 2
  • The Green Score (GS) expresses useful computational work per unit of real environmental impact, incorporating data-center PUE and per-node hardware efficiency for CPU-only or CPU+GPU profiles.Source: §1 · p. 2
  • Both metrics become green-affinity features in the AI4EOSC and DIRAC Green Directors while leaving existing scheduling logic unchanged.Source: §1 · p. 2
  • AI4EOSC cluster filling prioritizes greener providers without degrading latency or end-user experience; DIRAC evaluation reports 42.8% lower carbon emissions in simulation and 49.5% better carbon efficiency in production.The DIRAC evaluation covers a trace-driven simulation of 133,631 jobs and a preliminary KM3NeT deployment, while making the carbon-water trade-off explicit.Source: §1 · p. 2

2. Related work

Prior work addresses carbon-aware scheduling and increasingly recognizes water and other environmental impacts, but practical unified metrics for operational placement remain limited. GreenDirector combines cross-border electricity tracking and localized seasonal water stress in an actionable score used by federated schedulers.

2.1. From carbon-aware to water-aware computing

  • Carbon-aware scheduling research shifts workloads across locations or times to reduce emissions, while studies also report trade-offs involving spatial-temporal variability and hardware heterogeneity [14].Source: §2.1 · p. 2
  • Water-aware studies emphasize localized scarcity and water footprints, including scheduling approaches for geographically distributed AI workloads.Source: §2.1 · p. 2

2.2. Multi-dimensional environmental optimization

  • Multi-dimensional approaches consider carbon, water, and sometimes land impacts, while WaterWise shows that low-carbon power choices can worsen water footprints [22].Source: §2.2 · p. 3

2.3. Unified spatial-temporal metrics

  • Existing work lacks a practical standardized metric unifying environmental signals for operational placement, often relying on static water averages or omitting standardized scarcity factors.GreenDirector combines real-time cross-border electricity tracking with seasonal localized water-stress factors in an affinity score for federated orchestration.Source: §2.3 · p. 3

3. Methodology

The methodology traces electricity-related carbon and water impacts across regions and time, combines them in ES, and uses GS to rank workload placement by useful work per environmental impact. It applies these metrics to provider and node comparisons in AI4EOSC and DIRAC.

3.1. Impact metric calculation

  • Wattnet supplies high-resolution electricity carbon and water footprints, accounting for imports and exports; the method treats these operational electricity impacts as Scope-2-like and excludes on-site cooling.Carbon impact uses global CF because emissions have location- and season-independent climate effects, while water impacts depend on local and seasonal conditions.Source: §3.1 · p. 3

3.1.1. Water impact

  • AWARE2.0 factors vary seasonally across the studied zones, with Mediterranean AI4EOSC sites peaking above 70 while IISAS and PSNC stay below 6; DIRAC’s INFN-T1 reaches about 15.The observed variation supports using monthly rather than fixed annual factors.Source: §3.1.1 · p. 4, Figure 1 · p. 5
  • Global water impact weights each contributing generation zone’s electricity share and water-footprint intensity by its monthly AWARE2.0 scarcity factor.Monthly factors capture seasonal variation; spatial-temporal characterization is applied to operational electricity because its generation location and timing are known, not to other life-cycle stages.Source: §3.1.1 · p. 4

3.1.2. Unified environmental score

  • ES ratio-scales impact indicators against the most polluting available zone, applies fixed JRC-derived weights of 0.71 for CF and 0.29 for WI, and inverts the result into a 0–100 quality score.The score is zone-relative; ratio scaling does not force the cleanest zone to zero, and WI’s realized average contribution can be lower than 29%.Source: §3.1.2 · pp. 4–5
  • ES comparisons distinguish providers whose scores reflect water stress from those driven mainly by carbon intensity, including TUBITAK’s hydropower-related WI and the low WI of DIRAC sites.CC-IN2P3 has the highest ES among the nine providers at 97.4; PSNC is primarily penalized by carbon intensity.Source: §3.1.2 · p. 5, Figure 2 · p. 6

3.2. GreenScore

  • Hardware efficiency uses CPU and, for eligible AI workloads, GPU performance terms; workload profiles determine candidate providers and scaling subsets.For CPU+GPU workloads, the initial CPU and GPU energy shares are αC = 0.2 and αG = 0.8; AI4EOSC ranks eligible nodes using their hardware efficiency.Source: §3.2 · p. 6
  • GS multiplies hardware efficiency by PUE and the proportional environmental burden complement, representing useful computation delivered per unit of real environmental impact.Higher GS indicates better overall environmental efficiency and supports prioritizing greener execution sites.Source: §3.2 · pp. 5, 7
  • AI4EOSC’s CPU-only GS ranking favors IFCA, while PSNC’s stronger GPU hardware term does not overcome its grid impact; DIRAC’s site ordering stays stable across the observation period.The figures report monthly scores aggregated within each use case; their ratio-scaled hardware terms make comparisons across use cases invalid.Source: §3.2 · p. 7

4. Practical implementation

AI4EOSC adds green affinity after feasibility filtering, while DIRAC uses Green Score to order eligible sites for pilot submission. Both approaches retain their existing scheduling constraints and provisioning logic.

4.1. AI4EOSC allocation placement

  • AI4EOSC first filters for healthy, adequately resourced nodes and enforces readiness, namespace, and deployment-type constraints.Nodes must also meet the job’s CPU, GPU, RAM, disk, and GPU-model requirements.Source: §4.1 · pp. 7, 9
  • Nomad then ranks feasible nodes using bin packing and existing affinities, including GPU avoidance for CPU-only jobs, namespace preference, and soft green affinity.These preferences promote desired placements without blocking jobs when an affinity cannot be satisfied.Source: §4.1 · p. 9

4.1.1. AI4EOSC Green Director

  • The Green Director retrieves Wattnet ES, averages it over seven days, combines it with PUE and node hardware efficiency, then maps Green Scores to a softened 0–30 affinity.The initial score mapping is linear, with the green affinity scaled to remain softer than the other AI4EOSC affinities.Source: §4.1.1 · p. 9

4.1.2. Validation

  • AI4EOSC validates green affinity through score inspection and cluster filling, running the latter at maximum green affinity during a low-demand period.Both tests isolate the affinity by applying it only to one data center; deployments are deleted afterward.Source: §4.1.2 · p. 9

4.2.1. DIRAC GreenSiteDirector (GSD)

  • DIRAC’s GreenSiteDirector orders eligible data centers by decreasing Green Score before pilot submission, while retaining demand, compatibility, and capacity limits.Lower-ranked sites can still receive pilots when higher-ranked sites lack capacity or cannot serve the workload.Source: §4.2.1 · p. 10

4.2.2. Validation

  • Validation pairs a controlled trace replay that changes only site ordering with preliminary KM3NeT production observations, which are not a controlled policy comparison.The simulation uses fixed historical site characteristics; production results are subject to varying demand, site availability, queues, startup delays, failures, and compatibility.Source: §4.2.2 · pp. 11–12

5. Results and discussion

The results and discussion examine how GreenDirector’s carbon- and scarcity-weighted water signals affect workload placement and efficiency, alongside operational feasibility in AI4EOSC and DIRAC. They also describe a water-impact trade-off and limits on what GS’s proxy-based steering can guarantee.

5.1. Environmental characterization of electricity consumption by the infrastructure providers

  • In solar-cycle zones, shifting delay-tolerant jobs to 10:00–14:00 UTC can improve ES by up to 17 points, whereas baseload-dominated CC-IN2P3 offers essentially no time-shifting gain.The analysis uses 15-minute Wattnet data from July 2025 to June 2026, traces cross-border electricity exchanges, and covers operational electricity impacts rather than on-site cooling water.Source: §5.1 · pp. 12–13
  • ES combines the carbon gradient with scarcity-weighted water information: IISAS scores 80 versus IFCA’s 75 despite higher carbon intensity, while raw water footprint can mislead.Across zones, ES correlates −0.95 with CF and −0.14 with WI, whereas raw WF correlates +0.42 with ES; only IFCA and TUBITAK among the nine providers have appreciable water stress.Source: §5.1 · pp. 13–14

5.2. AI4EOSC GD validation

  • AI4EOSC validation showed that green affinity makes greener data-center nodes rank higher and fill first, reducing environmental impact without degrading end-user experience or platform requirements.An initially highly loaded IFCA node received the first 15 jobs because its binpack score outweighed green affinity; greener-site filling followed once that node filled.Source: §5.2 · pp. 14–15, Figure 9 · p. 15

5.3.1. Evaluation using trace-driven simulation with historically characterized sites

  • GS-based scheduling cut carbon emissions 42.76% but increased AWARE-weighted water-scarcity impact 55.64% versus randomized scheduling in the 133,631-job DIRAC replay.Carbon fell from 19.959 to 11.424 kgCO2e, while water-scarcity impact rose from 3.498 to 5.444 stress-L; prioritizing IN2P3-IRES explains the trade-off.Source: §5.3.1 · pp. 15, 17, Table 2 · p. 15
  • GS-based scheduling improved complete-run ES-derived environmental efficiency by 89.88%, delivering 156.10 versus 82.21 million normalized CPU-seconds per burden-weighted kWh.This combined-metric result remains favorable despite higher water-scarcity impact because carbon has the larger contribution under the selected ES weighting; physical carbon and water impacts are reported separately.Source: §5.3.1 · p. 17

5.3.2. Preliminary evaluation using DIRAC deployment in production

  • In preliminary KM3NeT production observations, GS ordering raised carbon efficiency 49.54%, while water-scarcity impacts were 2.532 versus 2.254 stress-L under random ordering.The periods were not a controlled comparison: pilot provisioning was opportunistic and operational and environmental conditions differed; participating sites had broadly similar, relatively low water-impact intensities.Source: §5.3.2 · pp. 17–18

5.4. Limitations

  • The evaluated ES covers carbon and water only, excludes on-site and embodied impacts, uses fixed relative weights, and can shift with the comparison set through ratio scaling.The authors note that the single index obscures the component trade-off; other electricity-generation externalities are omitted, and input assumptions and spatial-temporal resolution are taken as given.Source: §5.4 · p. 18
  • GS trades hardware efficiency against grid cleanliness using catalogue benchmarks and provisional CPU/GPU energy splits, so it can favor efficient sites on carbon-intensive grids.Hardware efficiency comes from nameplate benchmark scores, catalogue TDP and FP32 throughput, and only two workload profiles are distinguished.Source: §5.4 · p. 18
  • GS ranks impact intensity, not total impact, and placement cannot curb compute demand; its gains depend on federation heterogeneity and omit saturation and fairness.Disproportionate workload attraction can grow a site's absolute footprint; as grids converge, headroom shrinks, while load may concentrate until a provider fills without accounting for quality of service.Source: §5.4 · p. 18
  • Evaluation shows ES and GS steer placement, but does not independently measure avoided carbon or water, leaving proxy fidelity unverified.The evaluation assumes rather than demonstrates that the predictive affinities correspond to realised impact.Source: §5.4 · p. 18

6. Conclusion

GreenDirector integrates carbon and scarcity-weighted water impacts into workload placement, with evaluations in AI4EOSC and DIRAC showing greener placement and an explicit carbon-water trade-off. The approach can be extended with additional environmental impact categories.

6. Conclusion

  • The Environmental Score combines carbon footprint with spatial-temporal, stress-weighted water impact from electricity, using real-time cross-border tracking and monthly AWARE2.0 factors.It is a dimensionless index from 0 to 100.Source: §6 · p. 19
  • In AI4EOSC, ES affinities steered jobs toward greener providers without degrading end-user experience, scheduling latency, or hardware resource constraints.The evaluation covered four pan-European scientific cloud providers and found greener data centers filled first.Source: §6 · p. 19
  • 42.8% lower carbon emissions in simulation and 49.5% higher carbon efficiency in production came with higher water-scarcity impact when the lowest-carbon site had higher water stress.The simulated run also improved environmental efficiency by roughly 90% under the combined ES burden, which weights carbon more heavily under the chosen normalization.Source: §6 · p. 19
  • Further environmental impact categories can be added to the ES without changing per-site Green Score formulation or scheduling logic.The authors identify additional regional and local impact signals as future work.Source: §6 · p. 19

CRediT authorship contribution statement

The statement assigns conceptualization and supervision across several authors, while methodology, software, investigation, validation, writing, and project roles are distributed across the team.

CRediT authorship contribution statement

  • María Castrillo, Andrei Tsaregorodtsev, and Álvaro López García are credited with conceptualization; Castrillo, Tsaregorodtsev, and López García also held supervision roles.Tsaregorodtsev and López García additionally share project administration and funding acquisition.Source: CRediT authorship contribution statement · p. 19
  • Ignacio Heredia and Jaime Iglesias Blanco contributed methodology and software; Heredia also contributed investigation and validation, while Iglesias Blanco handled data curation and visualization.Both are credited with writing the original draft.Source: CRediT authorship contribution statement · p. 19
  • Castrillo, Tsaregorodtsev, Mazen Ezzeddine, and López García contributed methodology, investigation, and validation; Ezzeddine also contributed software, formal analysis, data curation, and visualization.Castrillo and Ezzeddine contributed to the original draft; Tsaregorodtsev and López García contributed to review and editing.Source: CRediT authorship contribution statement · p. 19

Data and code availability

Environmental indicators are obtained from Wattnet, and the AWARE2.0 characterization factors are published sources. The scheduling implementations are available in public code repositories.

Data and code availability

  • CF, WF, WI, and ES data at 15-minute resolution are obtained through Wattnet’s public API and are not redistributed with the work.The AWARE2.0 characterization factors are those published in.Source: Data and code availability · pp. 19–20
  • The DIRAC GreenSiteDirector and AI4EOSC GD implementations are available as open-source software through their respective project repositories.The DIRAC implementation extends DIRAC Interware under GPLv3; the AI4EOSC implementation is available under Apache-2.0.Source: Data and code availability · p. 20
Loading 2609.12602v2…