Power-Aware Scheduling for Multi-Center HPC Electricity Cost Optimization

TL;DR

Power-Aware Scheduling for Multi-Center HPC Electricity Cost Optimization tackles the challenge of rising HPC energy costs under dynamic pricing. It presents TARDIS, a framework that integrates a Graph Neural Network for fine-grained per-job power prediction with a multi-objective, spatial-temporal scheduling mechanism to minimize cost across multiple centers. Key contributions include a GNN-based power predictor, a scalable scoring function that combines cost, efficiency, and fairness, and a dispatch algorithm that leverages cross-site price heterogeneity. The approach yields substantial cost savings—up to 18% for temporal optimization and 10–20% in multi-site deployments—while preserving system throughput, demonstrating practical impact for sustainable, cost-efficient HPC operations.

Abstract

This paper introduces TARDIS (Temporal Allocation for Resource Distribution using Intelligent Scheduling), a novel power-aware job scheduler for High-Performance Computing (HPC) systems that minimizes electricity costs through both temporal and spatial optimization. Our approach addresses the growing concerns of energy consumption in HPC centers, where electricity expenses constitute a substantial portion of operational costs and have a significant financial impact. TARDIS employs a Graph Neural Network (GNN) to accurately predict individual job power consumption, then uses these predictions to strategically schedule jobs across multiple HPC facilities based on time-varying electricity prices. The system integrates both temporal scheduling, shifting power-intensive workloads to off-peak hours, and spatial scheduling, distributing jobs across geographically dispersed centers with different pricing schemes. We evaluate TARDIS using trace-based simulations from real HPC workloads, demonstrating cost reductions of up to 18% in temporal optimization scenarios and 10 to 20% in multi-site environments compared to state-of-the-art scheduling approaches, while maintaining comparable system performance and job throughput. Our comprehensive evaluation shows that TARDIS effectively addresses limitations in existing power-aware scheduling approaches by combining accurate power prediction with holistic spatial-temporal optimization, providing a scalable solution for sustainable and cost-efficient HPC operations.

Figures (12)

• Figure 1: Frontier's power consumption and energy efficiency over March 2023 sun2024energy.
• Figure 2: HPC job submission patterns over time.
• Figure 3: Power consumption patterns across HPC components and HPC jobs.
• Figure 4: Analysis of job trends and power consumption across job exit states.
• Figure 5: Block diagram of $\mathsf{TARDIS}$.