Energy Efficient and Grid-Interactive Datacenters
Exploiting workload flexibility in time and space to design energy efficient and grid-responsive data centers
Energy is the biggest constraint on AI and cloud data centers. Electricity grids are struggling to keep pace with rising demand and with data center demand patterns. Moreover, the electricity that powers these data centers is not always clean, and its cost and availability change frequently. By 2030, demand growth is projected to outpace grid decarbonization, and global data center emissions could rise by as much as 4.2x (Maji et al., 2025). My research exposes energy as a schedulable resource and exploits the flexibility of when and where workloads run to make data centers responsive to the grid.
Energy Virtualization
Today’s energy systems hide the unreliability of renewable power inside hardware, so software cannot see or manage it. Ecovisor (Souza et al., 2023) virtualizes the energy system the way a hypervisor virtualizes compute. Ecovisor gives each application software-defined control over its own energy, and each application decides how to handle the intermittency of clean power.
Grid-Responsive Scheduling and Scaling
CarbonScaler (Hanafy et al., 2024) uses that visibility to chase cleaner electricity. It exploits the elasticity of cloud workloads, scaling them up during clean-energy windows and down during dirty ones, and it builds carbon signals directly into the autoscaling loop. It won Best Student Paper at ACM SIGMETRICS 2024.
Scaling compute isn’t the only lever. FM-CAC (Yang et al., 2026) adds a new dimension for battery-buffered edge AI. It treats the battery as an active temporal buffer that separates when energy is drawn from the grid from when it is consumed, and it jointly optimizes the software pipeline, the hardware operating point, and the battery schedule at each step.
The same elasticity ideas also apply to cluster-level provisioning across heterogeneous workloads (CarbonFlex (Hanafy et al., 2025)), jobs with unknown length (LACS (Bostandoost et al., 2024)), the choice of the best scheduling policy per job (FTL (Bostandoost et al., 2024)), and workloads with internal task dependencies (job-shop-based carbon scheduling (Bostandoost et al., 2026)).
Spatial Shifting
Grid-responsiveness is also about where a workload runs. CarbonEdge (Wu et al., 2025) shows that carbon intensity varies significantly even at mesoscales, within a US state or among neighboring European countries, well below the continental scale that most cloud-region shifting exploits.
The same idea scales to CDNs. CDN-Shifter (Murillo et al., 2024) combines spatial load shifting with capacity shifting across a large CDN’s edge sites.
The principle extends to video streaming, where the network path matters as much as server placement (Go with the Flow (Murillo et al., 2026)).
Carbon–Cost–Energy Tradeoffs
Grid-responsive scaling raises a cost problem. Shifting a job to a cleaner window leaves already-purchased reserved capacity idle, which raises the effective cost of the work that does run. GAIA (Hanafy et al., 2024) co-optimizes carbon, cost, and performance together, and it uses real-time carbon-intensity data to decide how much reserved capacity to spend now versus shift later.
The War of the Efficiencies (Hanafy et al., 2024) exposes a deeper tension. Once the grid exposes time-varying carbon intensity, the most energy-efficient job is not always the most carbon-efficient one, so optimizing for one can work against the other.
Untangling the Carbon-Cost Tradeoffs (Hanafy et al., 2026) goes beyond single-application tradeoffs and quantifies the stampede effect at scale, where many independent schedulers converge on the same low-carbon windows or regions. That convergence creates the contention carbon-aware scheduling was meant to avoid and undercuts both cost and carbon goals.
