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

The rapid proliferation of data centers has led to massive energy demand and carbon emissions, posing significant sustainability challenges. The proposed framework addresses these limitations by combining workload scheduling with an utilization-dependent exponential aging model to evaluate long-term carbon costs from server degradation, achieving up to 13.0% lower carbon emissions and 12.6% lower operational costs compared with benchmarks.

The gist

The paper proposes a comprehensive carbon life-cycle modeling framework for distributed data centers that jointly accounts for operational carbon emissions and hardware-aging-induced embodied carbon emissions, resulting in a reduction of 13.0% in total carbon emissions and 12.6% in operational costs relative to benchmarks.

How it works

The proposed solution is structured around a lifecycle model that integrates three coupled flows: electricity flow, workload flow, and degradation flow. This framework aims to jointly minimize economic costs and lifecycle carbon emissions by optimizing the control variables across geographically distributed data centers.

1. Electricity Flow Modeling: This involves modeling the electricity purchase cost from the grid, which is defined as a function of active power drawn from the grid, electricity price, and duration of each time slot:

f g i,t = λ i,t Pg i,t ∆t.

2. Degradation-Informed Embodied Carbon Cost Model: This model links workload allocation to physical server aging and embodied carbon cost. It defines the resource utilization level ui,k,t as the ratio of assigned workload to total available processing capacity:

ui k,t ≜ Wi k,t / (N A i,k S rate i)

The load-dependent physical wear is mapped using an exponential aging model:

Luse i,k,t+1 = Luse i,k,t + ηi,k exp(γi k ui k,t ∆t) t ∈ [z]

The instantaneous embodied carbon expenditure for a cluster is calculated as:

Eemb i k,t = C M i k∆li k,t (8), where the fractional lifetime consumption is calculated as:

∆li k,t = ηi k exp(γi kui k,t)∆t / Llife i k.

The total embodied carbon cost for data center i at time slot t is aggregated across all active servers and converted into a monetary cost via the carbon price αC:

f c,m i,t = αC X k∈Ki N A i,k,t Eemb i k,t (9).

How it works (Continued)

The total carbon-related financial objective for data center i at time slot t is established as the sum of operational and embodied carbon costs:

f c i,t = f c o i,t + f c m i,t (10).

How it works (Continued)

The long-term optimization problem P1 is transformed into a sequence of per-slot deterministic optimization problems P2 using an online Lyapunov optimization framework to decouple the inter-temporal constraints. This involves defining three queue-based state variables: Batch Workload Queue (S D), Thermal Virtual Queue (τ˜i,t), and Battery Virtual Queue (E˜i,t).

The Time-Varying Queue Shifting (TVQS) mechanism adjusts the battery virtual-queue shift parameter ε i,t based on real-time electricity prices and renewable generation:

ε i,t = -Emax b i θ i,t (35), where the operational incentive factor θ i,t is defined by:

θ i,t = min (1, βθ / (P r i t P r rated i + P r i t + λ max i − λi,tλ max i !))

The unified Lyapunov function L(Θt) is defined as:

L(Θt) = 1/2 X i∈I [α i(S D i,t) squared + δ iE˜ squared i,t + ν iτ˜ squared i,t] (36).

Based on the provided scientific paper, here are specific improvements that can be made to AI systems, categorized by the capability they will gain:

) Carbon-Aware, Lifecycle-Optimized Resource Orchestration System (CLOROS): This system will move beyond traditional energy efficiency by integrating hardware degradation directly into its scheduling logic.

The improved system will dynamically allocate workloads across geographically distributed data centers not just based on real-time operational costs or immediate energy prices, but also by predicting the long-term embodied carbon cost associated with accelerated hardware aging (utilization-dependent thermal stress).

) Predictive, Adaptive Battery Management System: This system will replace static battery scheduling with a mechanism that actively anticipates future electricity price and renewable generation fluctuations.

The improved system will utilize the Time-Varying Queue Shifting (TVQS) mechanism to proactively charge or discharge energy storage based on forecasted electricity prices and renewable availability, leading to significantly more optimized energy arbitrage and reduced operational costs.

) Privacy-Preserving Distributed Workload Migration Network: This system will enable high-efficiency workload balancing across multiple independent data centers without compromising the security or privacy of local user/application data.

The improved system will employ the Zero-Sum Perturbation ADMM (ZSP-ADMM) framework to coordinate workload migration decisions across DCs in a way that masks individual migration volumes, ensuring global load balance is maintained while preventing adversaries from inferring sensitive local workload profiles.

) Hardware Health-Aware Cluster Selection Engine: This system will optimize server placement based on the predicted lifespan of different hardware components.

The improved system will prioritize allocating workloads to server clusters within a data center that exhibit lower degradation rates (as quantified by utilization and aging factors), thereby extending the effective operational life of the physical hardware and drastically reducing future embodied carbon emissions from replacement.

) Joint Economic and Environmental Optimization Controller: This is a holistic controller designed to minimize both immediate financial expenditure and total lifecycle carbon footprint simultaneously.

The improved system will solve a single, unified optimization problem (P1/P3) that balances the instantaneous operational costs (energy, transfer fees, operational carbon) against the long-term embodied carbon costs derived from hardware wear and tear, resulting in a globally optimal low-carbon deployment strategy.

This improved AI system can perform the following specific actions:

1. Identify and migrate latency-critical interactive workloads (IWs) to data centers offering the lowest combination of immediate electricity cost and future hardware degradation risk.

2. Manage battery energy storage in real-time, charging during periods of low carbon intensity or low electricity prices, and discharging strategically during high-cost/low-renewable peak hours, adapting its strategy dynamically via TVQS.

3. Balance the total workload across a global network of data centers by coordinating migration decisions using ZSP-ADMM, ensuring that no single DC's local workload profile is exposed to external coordination mechanisms.

4. Select the optimal cluster within a data center for any given task based on its utilization level and inherent hardware aging coefficient, ensuring that high-utilization tasks are directed toward younger hardware to minimize accelerated degradation and subsequent embodied carbon liability.

The rapid proliferation of data centers (DCs), driven by cloud computing and artificial intelligence (AI), has led to massive energy demand and carbon emissions, posing significant sustainability challenges. Carbon-aware optimization in geographically distributed data centers has been widely studied. Most existing approaches mainly focus on operational carbon emissions from server usage. However, existing literature often ignores workload-induced thermal stress, which accelerates nonlinear hardware degradation. This leads to more frequent server replacements and ultimately increases embodied carbon emissions. To address these limitations, we propose a comprehensive carbon life-cycle modeling framework for distributed data centers. Apart from operational carbon emissions, this work combines workload scheduling with a utilization-dependent exponential aging model to evaluate long-term carbon costs from server degradation. In order to solve the proposed optimization model in an online and privacy-preserving manner, an enhanced Lyapunov framework with time-varying queue shifting (TVQS) is first introduced to handle system uncertainties. Then, a zero-sum perturbation-based alternating direction method of multipliers (ZSP-ADMM) framework is developed to enable distributed coordination across geographically separated data centers while protecting locally exchanged workload information. Simulation results demonstrate that the proposed approach achieves up to 13.0% lower carbon emissions and 12.6% lower operational costs compared with benchmarks.