PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.00605v1eess.SYcs.SY

TL;DR

Existing approaches often focus on operational carbon and do not explicitly model utilization-dependent hardware degradation in lifecycle accounting. The paper combines aging-aware carbon modeling with online scheduling and privacy-preserving distributed coordination. It reports 13.0% lower total carbon emissions and 12.6% lower operational costs than the benchmark.

  • Problem

    Many existing approaches overlook utilization-dependent hardware degradation in lifecycle carbon accounting, while an existing model also relies on future system information and centralized coordination that raises privacy concerns.

    Source: §I · pp. 1–2

  • Method

    The framework combines a utilization-linked lifecycle carbon model with Lyapunov online scheduling using TVQS and ZSP-ADMM for distributed coordination.

    Source: §I · p. 2

  • Results

    13.0% lower total carbon emissions and 12.6% lower operational costs relative to the benchmark.

    Source: §I · p. 2

  • Takeaways & Limitations

    Combining workload migration with aging-aware scheduling reduces economic costs and lifecycle carbon emissions while preserving hardware health by prioritizing clusters with lower aging rates.

    Source: §V · p. 10

Abstract

from arXiv · show

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.

I. Introduction

The introduction identifies utilization-dependent hardware degradation as a missing part of lifecycle carbon accounting and proposes an online, distributed framework for geo-distributed data centers. The framework combines carbon and cost objectives with aging-aware scheduling, adaptive battery management, and workload coordination.

I. Introduction

  • Existing approaches largely omit utilization-dependent hardware degradation, while an earlier lifecycle model does not explicitly model it and relies on future information for scheduling.The paper notes that this can underestimate degradation-related embodied emissions; centralized coordination also raises privacy concerns.Source: §I · pp. 1–2
  • The proposed lifecycle model links server utilization to nonlinear lifetime degradation and reports 13.0% lower total carbon emissions and 12.6% lower operational costs than the benchmark.It jointly accounts for operational emissions and hardware-aging-induced embodied emissions.Source: §I · p. 2
  • A Lyapunov-based online method converts the joint economic-cost and lifecycle-carbon optimization into per-slot subproblems without future system information.TVQS adjusts the battery virtual-queue shift using real-time electricity prices and renewable generation.Source: §I · p. 2
  • ZSP-ADMM coordinates geographically separated data centers without directly exposing local workload migration decisions.Zero-sum perturbations preserve global workload balance while obfuscating local migration signals.Source: §I · p. 2

II. Proposed DC Operation Framework

  • The framework jointly optimizes electricity, workload, and degradation flows across data centers equipped with local renewables, storage, and heterogeneous server clusters.Workload allocation affects hardware degradation, which informs replacement cycles and embodied emissions.Source: §II · p. 2

A. Operational Cost Model

B. Degradation-Informed Embodied Carbon Cost Model

The model connects workload-driven server utilization to nonlinear aging and lifetime consumption, then amortizes manufacturing carbon according to the lifetime consumed. It also models interactive workload migration and local processing of batch workloads.

B. Degradation-Informed Embodied Carbon Cost Model

  • Higher server utilization accelerates physical aging, shortens server lifespan, and increases replacement-related embodied carbon.The model links workload utilization to server aging and the embodied carbon cost induced by hardware replacement.Source: B. Degradation-Informed Embodied… · p. 3
  • Manufacturing embodied carbon is amortized in proportion to lifetime consumed, so sustained high utilization increases carbon attributed to current operation.The model maps cluster utilization to lifetime consumption using an exponential aging model.Source: 2) Embodied Carbon Accounting · p. 3, 1) Utilization-induced Aging Model · p. 3

C. Workload Migration Model

D. Service Level Agreement (SLA) Constraint for Workload Processing

The workload-processing constraints approximate execution delay from server utilization and available computing capacity, and bound delay and active-server counts.

D. Service Level Agreement (SLA) Constraint for Workload Processing

E. Computing Power Consumption and Thermal Dynamics Model

The model relates cluster computing power to server power profiles and ambient temperature to thermal dynamics, while power-management constraints govern grid purchases, storage, and power balance.

E. Computing Power Consumption and Thermal Dynamics Model

F. Power Management Constraints

G. Joint Cost and Carbon Optimization Problem

The paper formulates joint cost-and-carbon optimization and decomposes it into online per-slot decisions. Its queue-shifting mechanism adapts battery incentives to current renewable generation and electricity prices.

G. Joint Cost and Carbon Optimization Problem

III. Solution Method

  • Lyapunov optimization decouples the long-term problem into per-slot subproblems without requiring forecasts of future renewable generation, electricity prices, or workload arrivals.Source: §III · p. 5

A. Unified Online Optimization Framework

1) Virtual Queue Modeling and Time-Varying Queue Shifting:

3) Per-slot Online Optimization Problem:

The per-slot problem is formulated for online decisions, while ZSP-ADMM coordinates migration without exposing exact local migration decisions. Coordinated zero-sum perturbations mask those decisions while preserving global workload balance.

3) Per-slot Online Optimization Problem:

B. Zero-Sum Perturbation ADMM Framework

  • ZSP-ADMM masks exchanged migration volumes by adding coordinated perturbations that sum to zero, leaving total migration volume unchanged.Sharing exact migration decisions would expose operational privacy risks.Source: B. Zero-Sum Perturbation ADMM Framework · p. 6
  • The perturbation scale is based on a conservative sensitivity bound: the maximum admissible change is set to Δm = Mmax.Preliminary perturbations are sampled from a zero-mean Laplace distribution with scale Δm/ϵ.Source: B. Zero-Sum Perturbation ADMM Framework · p. 6
  • ZSP-ADMM alternates local data-center updates with a global consensus update and local dual-multiplier updates.For fixed zero-sum perturbations, the reformulated problem is convex and has the same feasible physical decisions and optimal objective value as P2.Source: B. Zero-Sum Perturbation ADMM Framework · p. 7

IV. Numerical Results

The proposed online distributed algorithm is evaluated using real-world traces, with simulation setup and parameters detailed before its effectiveness is assessed.

IV. Numerical Results

  • The proposed online distributed algorithm is evaluated using real-world traces.The section first details the simulation setup and parameter configurations, then validates the proposed model and algorithm.Source: §IV · p. 7
  • The evaluation first details the simulation setup and parameter configurations.Source: §IV · p. 7
  • The section subsequently validates the effectiveness of the proposed model and algorithm.Source: §IV · p. 7

A. Parameter Setup

The simulation setup covers three data centers with distinct regional energy resources and heterogeneous server clusters. The online procedure uses observed system conditions to update scheduling and perturbation decisions.

A. Parameter Setup

  • The three data centers represent wind-dominated, hydro-based, and high-PV regions with different electricity-price and carbon-intensity profiles.Their profiles are shown in Fig. 3.Source: A. Parameter Setup · p. 7, Figure 3 · p. 8
  • At each time slot, the procedure observes workload, renewable generation, electricity price, carbon intensity, and system states before updating the TVQS shift.It also samples and coordinates perturbations before forming the per-slot optimization problem.Source: A. Parameter Setup · p. 8
  • The algorithm solves local subproblems, transmits obfuscated migration variables, and implements converged control decisions.System states are then updated for the next time slot.Source: A. Parameter Setup · p. 8
  • Each data center has four server clusters of 1,000 servers, distinguished by computing capacity and service age.Global simulation parameters are listed in Table II.Source: A. Parameter Setup · p. 8

B. Performance Comparison

The proposed method combines workload migration with aging-aware server allocation and outperforms the benchmark on cost and carbon emissions. Operational scheduling adapts battery use and server-cluster utilization to changing conditions, while perturbation masks migration signals.

B. Performance Comparison

  • Compared with the benchmark, M2 reduces total cost by 7.0% and carbon emissions by 8.0%, while M3 reduces total cost by 7.1% and carbon emissions by 7.7%.Source: B. Performance Comparison · p. 8

C. Operational Scheduling Analysis

B. Performance Comparison

  • 12.6% lower total cost and 13.0% lower carbon emissions than the benchmark; the proposed method records 23.47 ×10^3 $ and 148.04 t CO2, respectively.It jointly transfers workloads toward data centers with lower electricity prices and grid carbon intensities while prioritizing clusters with lower degradation rates.Source: Table III · p. 8, B. Performance Comparison · p. 8

C. Operational Scheduling Analysis

D. Impact of TVQS on Battery Scheduling

Appendix A Proof of the Lyapunov Drift Upper Bound

The appendix bounds the conditional one-slot Lyapunov drift by deriving bounds for individual queues and combining them. It also establishes that the per-slot optimization problem is convex by showing its objective is convex and its feasible region is defined by affine constraints.

Appendix A Proof of the Lyapunov Drift Upper Bound

  • The thermal virtual queue's squared transition has bounded higher-order and cross terms under the stated temperature bound.The proof identifies finite bounds for both the physical-dynamics terms and the control-independent cross-term.Source: Appendix A · p. 11
  • The conditional one-slot Lyapunov drift has a finite upper bound after the individual queue bounds are summed.The proof bounds higher-order terms and the terms introduced by the time-varying battery queue shift.Source: Appendix A · pp. 11–12

Appendix B Proof of the convexity of the problem

  • The exponential aging-amortized embodied carbon cost is jointly convex by the perspective construction, and the remaining objective terms are affine.The perspective is well defined when the server-capacity lower bound is positive.Source: Appendix B · p. 12
  • The per-slot feasible region is convex because the delay constraint and the remaining fixed-parameter constraints are affine.The proof also rewrites the utilization and computing-power relations as affine constraints.Source: Appendix B · p. 12

Appendix C Proof of the Equivalence Between P2 and P3

The appendix proves that P2 and P3 admit the same feasible physical decisions. Because corresponding feasible solutions also have identical objective values, the two problems have the same optimal objective value and corresponding optimal physical decisions.

Appendix C Proof of the Equivalence Between P2 and P3

  • Every feasible physical decision in P2 can be extended to a feasible solution of P3.The proof uses the zero-sum perturbation condition to show the remaining constraints are equivalent.Source: Appendix C · p. 12
  • Conversely, every feasible solution of P3 yields physical decisions feasible for P2, so the problems have the same feasible physical decisions.Together, the two directions establish equivalence of their physical feasible sets.Source: Appendix C · p. 12
  • Corresponding feasible solutions have identical objective values, giving P2 and P3 the same optimum and corresponding optimal physical decisions.The objective functions are identical for the same physical decisions and do not depend on zt.Source: Appendix C · p. 12
Loading 2610.00605v1…