⚡ electrical engineering

Aging-Aware Online Distributed Scheduling for Lifecycle Carbon Reduction in Geo-Distributed Data Centers

This paper proposes an aging-aware, online distributed scheduling framework that integrates operational and embodied carbon emissions through a utilization-dependent aging model and a privacy-preserving ZSP-ADMM algorithm, achieving significant reductions in both carbon footprint and operational costs for geo-distributed data centers.

Original authors: Junyu Lin, Wenjie Liu, Shunbo Lei, Wentian Lu, Jianhui Wang, Junhong Liu

Published 2026-10-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Junyu Lin, Wenjie Liu, Shunbo Lei, Wentian Lu, Jianhui Wang, Junhong Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The modern world runs on data centers, vast warehouses of computers that power everything from streaming videos to artificial intelligence. These facilities are hungry for electricity, and their growing appetite is driving up carbon emissions, the gases that trap heat in our atmosphere. For years, engineers have tried to make these centers greener by shifting their work to times and places where electricity comes from clean sources like wind and solar. However, this approach has a blind spot. It treats the computers inside the data centers as if they never wear out. In reality, running servers hard generates heat and stress, which speeds up their physical aging. When a server breaks down sooner than expected, it must be replaced. Manufacturing a new machine and disposing of the old one creates a massive amount of carbon, often more than the electricity the machine ever used. This hidden cost, known as embodied carbon, means that simply chasing cheap, clean electricity can sometimes backfire, causing more harm to the planet in the long run by forcing premature hardware replacement.

A team of researchers has developed a new way to manage these digital powerhouses that accounts for this wear and tear. They created a system that looks at the entire life of a server, not just its daily electricity bill. Instead of just asking where the cleanest power is available, their method also asks which computers are already tired and which are still fresh. By balancing the need for clean energy with the need to preserve hardware, they can schedule tasks to run on machines that are less likely to break, thereby reducing the total carbon footprint of the data center over its entire lifespan.

The researchers built a model that links how hard a server works to how quickly it degrades. They found that when a server is pushed to its limits, it heats up and ages faster, shortening its useful life. This accelerated aging means the center has to buy new equipment sooner, and the carbon cost of making that new equipment is added to the center's total environmental impact. To solve this, the team designed an online scheduling system. Unlike older methods that need to know the future to make plans, this system reacts in real-time. It watches the weather, the price of electricity, and the current load on the servers, then makes split-second decisions on where to send work. It uses a mathematical approach that treats the system's state like a balancing act, constantly adjusting to keep things stable without needing to predict tomorrow's weather or prices.

A major challenge in managing a network of data centers spread across different cities is privacy. Each center wants to keep its specific workload details secret from the others, yet they must coordinate to ensure the total amount of work stays balanced across the network. The researchers solved this with a clever communication trick. Instead of sharing their exact numbers, the centers exchange slightly altered versions of their data. They add random noise to the numbers they send, but they do it in a way that the noise cancels out when all the centers are added together. This allows the network to find the best global solution without any single center revealing its private operational secrets. The system ensures that the total work remains balanced while masking the specific details of what each center is doing.

When the team tested their approach using real-world data from three different locations, the results were clear. By combining the strategy of moving work to cleaner energy sources with the strategy of protecting hardware from unnecessary stress, they achieved significant improvements. Compared to standard methods that ignore hardware aging, their new system reduced total carbon emissions by 13.0 percent. It also cut operational costs by 12.6 percent. The simulations showed that the system successfully shifted work to servers that were less prone to rapid aging, extending their life and delaying the need for replacement. Furthermore, the system learned to charge batteries when electricity was cheap and abundant, and discharge them when prices were high, adapting its behavior to the changing conditions of the day.

The study also highlighted the importance of how batteries are managed. Older methods used a fixed rule for when to charge or discharge, which often led to conservative and inefficient use of the storage. The new system, however, adjusted its battery strategy based on the current price of electricity and the amount of renewable energy available. When the sun was shining or the wind was blowing, and prices were low, the system was more aggressive about charging. When conditions reversed, it released that stored energy. This flexibility allowed the data centers to make better use of clean energy without wasting storage capacity.

In the end, the research demonstrates that true sustainability in the digital age requires looking beyond the plug. It is not enough to just switch to green electricity; one must also care for the machines that consume it. By treating the physical wear of computers as a critical factor in scheduling, the researchers have shown a path toward data centers that are not only cheaper to run but also kinder to the planet over the long term. The findings suggest that a holistic view, which considers both the energy used and the materials wasted, is essential for the future of computing.

Technical Summary: Aging-Aware Online Distributed Scheduling for Lifecycle Carbon Reduction in Geo-Distributed Data Centers

Problem Statement
The rapid expansion of data centers (DCs) driven by cloud computing and AI has resulted in massive energy consumption and carbon emissions. While existing research has extensively addressed "operational carbon" (emissions from electricity usage), it frequently overlooks "embodied carbon" associated with hardware manufacturing and disposal. Specifically, current literature often neglects the impact of workload-induced thermal stress on hardware degradation. High server utilization accelerates physical wear, shortening server lifespans and necessitating more frequent replacements. This oversight leads to an underestimation of the total lifecycle carbon footprint, as the embodied carbon from accelerated hardware turnover is not factored into dynamic scheduling decisions. Furthermore, existing solutions often rely on offline optimization requiring future system knowledge or centralized coordination that compromises data privacy.

Methodology
The authors propose a comprehensive, online, and distributed framework to minimize both economic costs and total lifecycle carbon emissions (operational + embodied). The approach integrates three coupled flows: electricity, workload, and hardware degradation.

Lifecycle Carbon Modeling:
- Operational Carbon: Calculated based on grid electricity consumption and local carbon intensity.
- Embodied Carbon: A novel utilization-dependent exponential aging model is introduced. The model links server cluster utilization ( $u_{i,k,t}$ ) to a cumulative lifetime consumption metric ( $L^{use}_{i,k,t}$ ). High utilization accelerates aging, increasing the rate of lifetime consumption. The embodied carbon cost is amortized in real-time based on this accelerated consumption, converting the physical wear into a monetary cost via carbon pricing.
Online Optimization via Lyapunov Framework:
- To handle the uncertainty of future workloads, renewable generation, and electricity prices without requiring offline predictions, the long-term optimization problem is transformed into per-slot subproblems using Lyapunov optimization.
- Time-Varying Queue Shifting (TVQS): An enhanced mechanism is introduced to manage battery energy storage. Unlike standard methods using static virtual queue shifts, TVQS dynamically adjusts the battery virtual queue based on real-time renewable generation and electricity prices. This allows the system to adaptively incentivize charging during low-price/high-renewable periods and discharging during high-price periods.
Privacy-Preserving Distributed Coordination:
- To solve the optimization problem across geographically separated DCs without exposing local workload migration patterns, a Zero-Sum Perturbation-based Alternating Direction Method of Multipliers (ZSP-ADMM) is developed.
- Each DC adds a structured perturbation to its migration signals. These perturbations are coordinated to sum to zero globally, ensuring the total workload balance constraint is satisfied while obfuscating individual DC decisions. This preserves the global consensus structure while masking local operational data.

Key Contributions

Comprehensive Lifecycle Modeling: The paper develops a carbon model that jointly accounts for operational emissions and hardware-aging-induced embodied carbon. It mathematically links cluster-level utilization to nonlinear lifetime degradation, demonstrating that ignoring this factor underestimates lifecycle carbon.
Adaptive Online Algorithm: A Lyapunov-based online method is formulated that transforms the long-term stochastic problem into deterministic per-slot subproblems. The inclusion of the TVQS mechanism enables adaptive battery management responsive to real-time market and environmental conditions.
Privacy-Preserving Distributed Mechanism: The ZSP-ADMM framework enables distributed coordination across DCs. By injecting zero-sum perturbations into the exchanged migration variables, the method protects local workload privacy while maintaining global workload balance and achieving convergence with negligible computational overhead.

Results
Simulation results based on real-world traces from three interconnected DCs (with wind, hydro, and solar resources) compare the proposed method against benchmarks (no migration, no aging model, or partial implementations).

Performance: The proposed framework achieves a 13.0% reduction in total carbon emissions and a 12.6% reduction in operational costs compared to the benchmark.
Mechanism Efficacy:
- The aging-aware scheduling prioritizes server clusters with lower degradation rates, reducing the equivalent operating time of high-aging clusters by approximately 37–39% compared to benchmarks.
- The TVQS mechanism effectively shifts battery charging to periods of abundant renewable energy and low electricity prices, avoiding the conservative scheduling seen in fixed-shift methods.
- The ZSP-ADMM algorithm converges within 70 iterations, similar to standard ADMM, while successfully masking migration signals (Mean Absolute Error < 0.21 in normalized units) without disrupting global coordination.

Significance
The paper argues that effective carbon reduction in data centers requires a holistic view of the hardware lifecycle, not just operational energy efficiency. By integrating hardware degradation into the scheduling loop, the proposed framework prevents the "hidden" carbon costs of accelerated hardware replacement. Furthermore, the combination of online adaptability and privacy-preserving distributed coordination addresses critical barriers to the industrial deployment of carbon-aware scheduling in multi-tenant, geographically distributed environments. The study demonstrates that balancing economic efficiency with hardware longevity is essential for achieving true sustainability in the era of AI and cloud computing.

More like this