DataCenterGym: Physics-Grounded Simulator for Multi-Objective Data Center Scheduling
DataCenterGym has launched a simulation environment rooted in physics, aimed at optimizing job scheduling in geo-distributed data centers, which will serve as a reusable platform for upcoming research. This simulator encompasses compute queuing, thermal dynamics of buildings, localized HVAC operations, and service degradation influenced by temperature, all within a Gymnasium-compatible framework. Modern data centers manage diverse workloads across various locations, each with unique compute capabilities, electricity costs, and thermal conditions. The interplay between compute utilization, heat output, cooling requirements, and energy use is significant, yet many current schedulers overlook these interdependencies. A Hierarchical Model Predictive Control (H-MPC) scheduling algorithm has been created to facilitate distributed job placement while considering thermal and power dynamics. Tests were performed on standard operations and workload sensitivity.
Key facts
- DataCenterGym is a physics-grounded simulation environment for job scheduling in geo-distributed data centers
- The simulator integrates compute queueing, building thermal dynamics, localized HVAC behavior, and temperature-dependent service degradation
- It provides a Gymnasium-compatible interface
- 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
- Most existing schedulers abstract these effects and treat them independently
- A Hierarchical Model Predictive Control (H-MPC) scheduling algorithm was developed
- The algorithm performs distributed job placement while explicitly accounting for thermal and power dynamics
Entities
—