Towards Joint Energy Efficiency and Performance Optimization: A Scheduling Strategy for Green-Aware Heterogeneous Computing

Abstract

Against the backdrop of unprecedented growth in global computing demands and the rapid expansion of the artificial intelligence (AI) industry, the energy consumption and carbon emissions associated with computing infrastructure have emerged as pivotal challenges, significantly impeding the green transformation of the digital economy. The escalating reliance on high-performance computing (HPC) systems, data centers, and edge computing devices to support AI-driven applications, big data analytics, and cloud services has led to a substantial increase in energy use, exacerbating environmental concerns. This surge in energy demand not only strains power grids but also contributes to greenhouse gas emissions, undermining sustainability efforts in the digital sector. A critical challenge in this context is the spatio-temporal mismatch between resource scheduling and green energy supply within heterogeneous computing power networks. These networks, which integrate diverse computing resources such as CPUs, GPUs, and specialized accelerators, often operate in environments where renewable energy sources like solar and wind are intermittent. The variability in renewable energy generation, coupled with the dynamic nature of computing workloads, creates a complex scheduling problem. Traditional resource allocation strategies, which predominantly prioritize performance metrics such as latency and throughput, fail to account for the availability and utilization of green energy. Consequently, computing nodes frequently rely on non-renewable energy sources, leading to higher carbon footprints and inefficiencies in energy use. To address these challenges, this paper proposes a green energy-aware computing scheduling strategy designed to optimize the spatio-temporal coordination between heterogeneous computing resources and renewable energy supply. The strategy is grounded in a comprehensive sensing framework that integrates three key dimensions: task latency sensitivity, computational preference characteristics, and node-level green energy reserves. By analyzing task latency sensitivity, the strategy identifies workloads that can tolerate flexible execution times, enabling them to be scheduled during periods of abundant renewable energy. Computational preference characteristics, which include the type of processing required (e.g., CPU-intensive or GPU-intensive tasks), are used to match tasks with the most suitable computing nodes, thereby enhancing overall efficiency. Node-level green energy reserves are continuously monitored to ensure that tasks are allocated to nodes with sufficient renewable energy availability, minimizing reliance on non-renewable sources.