Multi-Objective Deep Reinforcement Learning for Carbon-Aware Spatiotemporal Workload Scheduling in Geo-Distributed Data Centers
Bibliographic record
Abstract
The rapid expansion of artificial intelligence training and cloud computing workloads has transformed United States data centers into major contributors to national carbon emissions, consuming between 1% and 1.3% of total national electricity output with projections indicating sustained double-digit annual growth. A fundamental yet underexploited characteristic of the US power grid is the spatiotemporal heterogeneity of carbon intensity: marginal emission rates vary by a factor of 5–10 across the seven major independent system operator (ISO) regions and exhibit pronounced diurnal and seasonal oscillations driven by renewable penetration patterns. Existing scheduling frameworks optimize for throughput and operational cost while treating carbon emissions as an externality, leaving substantial decarbonization potential untapped. This paper presents a multi-objective deep reinforcement learning (MO-DRL) framework that jointly exploits temporal deferral and geographic migration to minimize carbon emissions, job completion latency, and operational cost for delay-tolerant batch workloads across geo-distributed data centers. By formulating the scheduling problem as a multi-objective Markov decision process (MDP) and training a Pareto-conditioned policy network using multi-objective proximal policy optimization (MO-PPO), the proposed approach learns a rich set of Pareto-optimal scheduling strategies that enable operators to navigate the three-way tradeoff without rerunning optimization. Evaluated against real carbon intensity traces from six US ISO regions and Google/Alibaba cluster workload datasets, the framework achieves up to 41.3% carbon reduction compared to carbon-agnostic baselines while maintaining 95th-percentile job completion time within a 15% overhead bound.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.005 | 0.001 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.000 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.001 | 0.000 |
| Open science | 0.004 | 0.003 |
| Research integrity | 0.000 | 0.002 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".