Carbon-Aware Computing in a Network of Data Centers: A Hierarchical Game-Theoretic Approach

TL;DR

The work addresses carbon-aware computing across a global network of data centers by formulating a single-leader multi-follower Stackelberg game in which the DC operator sets virtual capacity curves to influence how flexible batch jobs are allocated in space and time. A hyper-gradient–based BIG Hype algorithm solves the bilevel problem by leveraging the equilibrium mapping of the lower-level game and its sensitivity to the leader's decisions. Numerical experiments with real carbon-intensity data show that temporal shifting and spatial migration can substantially reduce emissions, and co-designing VCCs with allocations yields more balanced waiting times than scheduler-agnostic approaches. Compared to sequential optimization, the bilevel approach achieves comparable decarbonization and peak shaving at high migration costs while enhancing fairness and responsiveness, highlighting its practical value for coordinated carbon-aware DC operation.

Abstract

Over the past decade, the continuous surge in cloud computing demand has intensified data center workloads, leading to significant carbon emissions and driving the need for improving their efficiency and sustainability. This paper focuses on the optimal allocation problem of batch compute loads with temporal and spatial flexibility across a global network of data centers. We propose a bilevel game-theoretic solution approach that captures the inherent hierarchical relationship between supervisory control objectives, such as carbon reduction and peak shaving, and operational objectives, such as priority-aware scheduling. Numerical simulations with real carbon intensity data demonstrate that the proposed approach successfully reduces carbon emissions while simultaneously ensuring operational reliability and priority-aware scheduling.

Figures (6)

• Figure 1: DC network featuring 4 DC locations interconnected by physical connection lines (solid line). A job migration between DC 1 and DC 3 is realized via a path through the fiber network. Right: Teams are associated with a DC location, where they initially submit their compute jobs.
• Figure 2: An example of sequential allocation of two compute jobs $y^1$ and $y^2$ on a data center $d \in \mathcal{D}$, where job 1 is of higher priority than job 2. The virtual capacity curve (red dashed line) limits the allocable load at each time slot. The maximum capacity of the DC $x^{\max}_{d,t}$ (solid black line) is obtained by subtracting the inflexible load from the maximum capacity of that DC $x^{\max}_d$ (dashed black line).
• Figure 3: Carbon emission savings normalized by compute volume due to time-shifting. Bilevel game vs, naïve approach (full capacity utilization) in three scenarios.
• Figure 4: Carbon emissions, normalized by compute volume, vs migration price $\xi$, for three scenarios. The growth demonstrates a positive impact of spatial migration.
• Figure 5: Fairness and total waiting time in a direct comparison between our approach (BL) and sequential optimization (SQ). Sequential optimization scores higher (worse) values in both, and the difference grows with increasing migration price.