DataCenterGym: A Physics-Grounded Simulator for Multi-Objective Data Center Scheduling

Abstract: Modern datacenters schedule heterogeneous workloads across geo-distributed sites with diverse compute capacities, electricity prices, and thermal conditions. Compute utilization, heat generation, cooling demand, and energy consumption are tightly coupled, yet most existing schedulers abstract these effects and treat them independently. We present \textit{DataCenterGym}, a physics-grounded simulation environment for job scheduling in geo-distributed data centers, designed as a reusable testbed for future research. The simulator integrates compute queueing, building thermal dynamics, localized HVAC behavior, and temperature-dependent service degradation within a Gymnasium-compatible interface. We also develop a Hierarchical Model Predictive Control (H-MPC) scheduling algorithm that performs distributed job placement while explicitly accounting for thermal and power dynamics. Through experiments on nominal operation and workload sensitivity, we demonstrate how H-MPC improves scheduling performance relative to baseline schedulers.

• The paper introduces a physics-grounded simulator that couples job scheduling with thermal and power dynamics for multi-objective data center optimization.
• It employs hierarchical MPC to dynamically adjust scheduling and cooling, achieving lower energy per job and reduced queue backlogs.
• Experimental results demonstrate that DataCenterGym reliably maintains thermal safety and efficiency under varying workloads.

DataCenterGym: Physics-Grounded Simulation for Multi-Objective Data Center Scheduling

Introduction and Motivation

The expansion of hyperscale and edge data centers has intensified challenges in managing compute utilization, energy consumption, and thermal safety, especially across geodistributed sites with heterogeneous climates, power limits, and hardware profiles. Existing scheduling infrastructures typically decouple job allocation from physical phenomena—such as heat generation, cooling demand, and power dynamics—leading to inefficiencies and operational risk. Most current simulators and schedulers lack closed-loop, physics-based feedback, making it difficult to evaluate policies that anticipate and mitigate cascading effects from workload placement to thermal stress and energy cost.

DataCenterGym directly addresses this gap by offering a modular, physics-driven simulation environment that couples job scheduling decisions with underlying thermal and power dynamics, modeled in a Gymnasium-compatible API. The environment aims to facilitate the development and rigorous evaluation of advanced scheduling policies that jointly optimize throughput, latency, energy, and thermal safety.

Architecture and System Dynamics

The core novelty of DataCenterGym lies in its integration of multiple subsystems:

• Workload and Queueing Model: Jobs drawn from real-world traces (Alibaba 2018) are annotated with resource demands, execution duration, priority, and hardware affinity, ensuring production realism in scheduling constraints.
• Thermal Modeling: Data centers are abstracted using a lumped RC thermal circuit, coupling compute workload to internal temperature via heat generation coefficients and modeling both passive and active dissipation via cooling systems.
• Power Management: Electrical capacity, cooling power, and their interdependence are explicitly simulated, with locational and temporal variability in grid prices.
• Closed-Loop Control: Scheduling, admission, and cooling setpoints are manipulated in concert (Figure 1), simulating physical feedback on multiple timescales.

  Figure 1: Closed-loop interaction in DataCenterGym. The scheduler observes the system state, selects job assignment and cooling actions, and the environment advances via coupled workload execution, thermal dynamics, and power evolution.

The simulator exploits a hierarchical decomposition, allowing tractable assignment and control across multiple data centers and clusters, while preserving realistic temporal lags—thermal inertia and delayed cooling effects.

Scheduling Policy Landscape

A comprehensive suite of policies is instantiated within DataCenterGym:

• Random and Greedy: Provide lower and utilization-optimizing baselines, respectively.
• Thermal-Aware and Power-Cool Heuristics: Use myopic proxies for thermal headroom and instantaneous energy cost, but lack anticipation of cascading effects.
• Safety-Constrained MPC (SC-MPC): Enforces hard limits on thermal and capacity state; computationally challenging beyond simple settings.
• Hierarchical MPC (H-MPC): Decomposes control into (i) a supervisory data center-level MPC for admission and thermal planning, and (ii) per-datacenter, cluster-level scheduling MPCs, enabling scalable and anticipatory joint optimization.

Experimental Results

Nominal Operating Regime

Under typical loads (target utilization 60–70%), H-MPC distinguishes itself from baselines:

• Queue Regulation: H-MPC minimizes queue backlogs—CPU queue and GPU queue lengths reduced relative to all heuristics.
• Utilization/Throughput: Rather than saturating clusters, H-MPC maintains a utilization target, preventing congestion and instability. Greedy and SC-MPC policies push toward saturation, but do not improve queueing delay.
• Thermal Safety: All policies maintain safe thermal margins, but SC-MPC is conservative, increasing operational cost. H-MPC tunes setpoints dynamically, minimizing unnecessary cooling.
• Energy Efficiency: H-MPC achieves the lowest energy per completed job (2.20 kWh/job) and overall operational cost (\$14,424 per 24h), outperforming all baselines on both throughput and cost objectives.

Sensitivity to Workload Intensity

As job arrival rate $\lambda$ increases, system behaviors diverge significantly across policies:

• Saturation Transition: Greedy and PowerCool policies experience a sharp knee in utilization and queue length just beyond $\lambda \approx 1.6\times$, corresponding to transition into congestion-dominated operation (Figure 2).
• Thermal Stress: Beyond this point, Greedy causes thermal excursions that activate throttling, limiting throughput.
• Anticipatory Control with H-MPC: H-MPC continuously biases system operation toward the nominal (safe and efficient) regime, extending the safe operational envelope and delaying entry into the saturation region.

Figure 3: Thermal response under increasing workload; H-MPC tightly regulates temperatures while Greedy incurs overheating and PowerCool eventually loses control as load increases.

Figure 2: System saturation under increasing load; H-MPC maintains target utilization and queue stability while Greedy/PowerCool policies track the plant's intrinsic saturation frontier.

Key Numerical Results:

• No Throttling in Nominal Regime: All policies avoid throttling under baseline load, but only H-MPC continues to do so as arrival rate increases.
• Energy/Job under Load: H-MPC maintains 2.20–2.35 kWh/job even as load increases, while heuristics degrade rapidly.

Modeling Assumptions and Limitations

• Thermal Granularity: Uses a lumped RC model at the datacenter level; does not model spatial gradients or per-machine/junction temperature.
• Network Abstraction: Ignores network latency, bandwidth, and data locality, which may influence effective scheduling in WAN contexts.
• Queue and Service Model: FIFO with backfilling; no consideration of dependencies or fragmentation.
• Fixed Dynamics: Power supply is exogenously modeled; coupling with renewables or grid demand response omitted.
• Learning-Based Control: Current version is model-driven; integrating RL or hybrid learning-MPC is left for future work.

Implications and Future Directions

Practical implications of DataCenterGym are significant:

• Framework for Multi-Objective Control Research: Serves as a testbed for queue/throughput/energy/thermal multi-objective optimization methods, including scalable RL, model-based control, and hybrid algorithms.
• Infrastructure-Aware Scheduling: Promotes the adoption of policies that explicitly trade off short-term performance for long-term reliability and cost—crucial for sustainable datacenter operations.
• Benchmarking Platform: Provides a reproducible simulation platform for comparative evaluation of new scheduling algorithms under realistic closed-loop physical constraints.

Theoretically, the decomposition of hierarchical MPC for mixed-integer, hybrid-action scheduling in cyber-physical systems offers a pathway toward scalable, safe, and anticipatory resource management. The approach also sets the stage for integration with learning-based model predictive architectures that can accommodate non-stationary operating conditions and model mis-specification.

Conclusion

DataCenterGym provides the research community with a rigorous, physics-grounded simulation environment for the study of modern data center scheduling. By bridging the gap between job placement models and underlying thermal and power dynamics, it enables principled exploration of trade-offs in multi-objective scheduling and paves the way for more sustainable and robust data center operations. The hierarchical MPC approach introduced demonstrates significant gains in cost, queueing, and thermal robustness, particularly as system load approaches saturation. Future work integrating learning-based methods and finer-grained physical modeling will further enhance the realism and applicability of the platform.

Reference:

"DataCenterGym: A Physics-Grounded Simulator for Multi-Objective Data Center Scheduling" (2604.15594)