What question did this study set out to answer?

To develop a carbon-aware job scheduling framework that minimizes energy costs and carbon footprints under uncertainty in cloud data centers.

June 7, 2026Journal of Cloud Computing Advances Systems and Applications1 citationsOpen Access

Carbon-aware scheduling and distributionally robust optimization for cloud systems: forecasting, battery storage, and adaptive ambiguity refinement

HRHardik RuparelFilm Independent

Key Points

To develop a carbon-aware job scheduling framework that minimizes energy costs and carbon footprints under uncertainty in cloud data centers.
Introduced a DRO framework incorporating LSTM-based forecasting for price and carbon-intensity predictions.
Developed an adaptive ambiguity-set refinement mechanism to manage forecast errors in real-time.
Evaluated over a benchmark of six scheduling methods against one thousand out-of-distribution scenarios.
DRO scheduler reduced CVaR carbon emissions to 103.6 kg, a 13.6% reduction compared to the cost-only baseline of 119.9 kg.
Achieved near-lowest expected carbon emissions of 89.2 kg, slightly above the carbon-first policy's 89.1 kg.
Maintained zero SLA violation rates across all methods, ensuring prompt service delivery.

Abstract

Abstract Geo-distributed cloud data centers consume vast amounts of energy and contribute substantially to carbon emissions, creating pressing economic and environmental concerns. This paper presents an extended and substantially enhanced distributionally robust optimization (DRO) framework for carbon-aware job scheduling under uncertainty. Building on a multi-objective formulation that balances energy cost, carbon footprint, latency penalties, and tail-risk control, the extended framework integrates four new technical components: (i) a data-driven lstm -based forecasting module that generates per-region price, carbon-intensity, and renewable-availability forecasts from historical traces, with an exponentially-weighted moving-average (EWMA) fallback; (ii) an adaptive ambiguity-set refinement mechanism that tracks forecast residuals online and shrinks or expands the Wasserstein radius in response to observed prediction errors; (iii) a battery-storage dispatch model whose charge/discharge dynamics are formulated as linear constraints compatible with the DRO framework, allowing energy to be stored during low-carbon periods and released during high-carbon periods; and (iv) a broad six-method comparative benchmark—including cost-minimizing, carbon-first, renewable-first, score-based, robust, and DRO schedulers—evaluated over one thousand out-of-distribution stress scenarios. Theoretical analysis establishes convexity of all subproblems and finite convergence of the Benders decomposition. On the large-scale comparative benchmark, the proposed DRO scheduler achieves the lowest CVaR ₀. ₉ carbon (103. 6 kg versus 119. 9 kg for the cost-only baseline, a 13. 6% reduction) and near-lowest expected carbon (89. 2 kg, marginally above the 89. 1 kg of the carbon-first policy) while incurring only a 13. 6% cost premium relative to the cost-minimizing baseline (28. 17 versus 24. 79 for CM and 31. 64 for the carbon-first policy) —demonstrating that DRO achieves near carbon-first sustainability at substantially lower cost than dedicated carbon-minimizing approaches. Battery storage experiments demonstrate temporal carbon arbitrage, with data centers charging during low-carbon windows and discharging during high-carbon peaks, effectively decoupling computation timing from grid carbon intensity. All Service Level Agreement (SLA) violation rates are identically zero across methods, confirming that hard deadline guarantees are maintained. The framework provides a rigorous, empirically validated foundation for sustainability-driven and risk-averse orchestration of cloud and distributed systems under deep uncertainty.

KI fragen

Bookmark

View Full Paper