Research map / Multi-agent systems and games / LLM agents

DCcluster-Opt: Benchmarking Dynamic Multi-Objective Optimization for Geo-Distributed Data Center Workloads

NeurIPS 2025 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces DCcluster-Opt, an open-source Gymnasium benchmark for carbon-, cost-, SLA- and water-aware geo-temporal AI workload scheduling across geo-distributed data centers, with rule-based and SAC/RLlib baselines and an LLM distillation case study.

Problem

Geo-distributed AI workload scheduling for sustainability requires realistic testbeds that jointly model time-varying grid carbon, electricity prices, weather, data center physics (IT, HVAC, water), and network cost/latency; existing benchmarks lack this integration.

Approach

A high-fidelity, physics-informed simulation benchmark that combines real-world geo-temporal datasets (Alibaba GPU trace, carbon intensity, electricity prices, weather, cloud transmission costs, empirical network delays) with a modular multi-objective reward and a Gymnasium API, plus baseline controllers and an LLM policy-distillation case study.

Key results

• SAC Geo+Time achieved $92,401 total cost and 308.7 t CO2 with 25.47% SLA violation, versus Round Robin's 0% SLA violation at $100,896 cost.
• Adding RL-controlled HVAC to SAC Geo+Time reduced total energy by 11.2% (to 921.3 MWh) and CO2 by 11.5% (to 273.1 t) with a 12.0% cost saving.
• The distilled LLM controller achieved 22.51 t carbon and 1.15% SLA on the 3-DC setting, slightly better than the expert SAC (22.91 t, 1.56%).

Measured effects reported in the paper

Surprising findings

• The SAC Geo+Time agent achieves lower cost and CO2 than most rule-based controllers but at a much higher SLA violation rate (25.47%) due to deferral.
• PPO agents trained with RLlib default hyperparameters performed worse than the custom SAC on cost, SLA violations and deferrals, indicating sensitivity to tuning.

Limitations

• Workload driven by a 2020 Alibaba GPU trace rather than modern LLM traces.
• Physics models are simplified; does not model locational marginal pricing or reactive electricity pricing in current experiments.
• High SLA violation rates for the best-cost RL agents show the reward/deferral design still needs tuning.

Experimental setup

Datasets: Alibaba Cluster Trace 2020 Electricity Maps carbon intensity GridStatus electricity prices Open-Meteo weather cloud provider transmission costs Persico et al. empirical network delay data

Benchmarks: DCcluster-Opt simulated 5-DC cluster 30-day evaluation 3-DC in-distribution and 5/10-DC out-of-distribution LLM tests

Metrics: total cost ($) total CO2 (t) total energy (MWh) total water (m3) SLA violation rate (%) average CPU/GPU utilization (%) transmission cost ($) tasks deferred delay (min)

Baselines: RBC Local Only RBC Lowest Carbon RBC Lowest Price RBC Most Available RBC Round Robin PPO APPO IMPALA

Builds on: SustainDC Soft Actor-Critic Ray RLlib Alibaba cluster trace 2020 Gymnasium

Closest papers by meaning

• Carbon Footprint Reduction for Sustainable Data Centers in Real-Time (AAAI 2024 · similarity 0.90)
• Hierarchical Multi-Agent Framework for Carbon-Efficient Liquid-Cooled Data Center Clusters (AAAI 2025 · similarity 0.90)
• SustainDC: Benchmarking for Sustainable Data Center Control (NeurIPS 2024 · similarity 0.89)
• LC-Opt: Benchmarking Reinforcement Learning and Agentic AI for End-to-End Liquid Cooling Optimization in Data Centers (NeurIPS 2025 · similarity 0.85)
• Data Center Cooling System Optimization Using Offline Reinforcement Learning (ICLR 2025 · similarity 0.85)
• Sustainability of Data Center Digital Twins with Reinforcement Learning (AAAI 2024 · similarity 0.85)