Flex: Grid-Responsive Provisioning and Scheduling for Elastic Cloud Clusters

Abstract

Computing energy demand and grid capacity growth are at a crossroads. Data centers are becoming large, concentrated electricity consumers just as power grids face capacity shortages while also integrating more variable renewable generation, causing grid conditions to change significantly across hours and seasons. These variations create an opportunity for \emph{grid-responsive} operations, in which workloads are shifted to align energy use with time-varying grid conditions, prioritizing clean, low-cost periods of surplus generation and low congestion. Fortunately, a significant fraction of computing demand comes from batch jobs that are delay-tolerant and elastic, enabling clusters to shift or scale jobs, or adjust the provisioned capacity in response to grid conditions. However, provisioning and scheduling decisions are closely coupled. Effective cluster capacity depends on how jobs are scaled, while job-scaling decisions depend on the available capacity. At the same time, exogenous grid signals (\eg carbon intensity and electricity prices) determine the cost of both scheduling and provisioning decisions. To address this problem, we present \Flex, a grid-responsive resource manager for elastic cloud clusters that coordinates provisioning of cluster capacity and scheduling of elastic batch jobs. Flex takes a data-driven approach, periodically learning optimal scheduling and provisioning decisions based on recent history and reusing them to guide coherent runtime provisioning and scheduling. We implement Flex on top of AWS ParallelCluster and evaluate it using real-world workload traces and grid signals, including carbon intensity and electricity prices. Our results show that Flex reduces carbon emissions by up to $\sim$57\%, within 2.1\% of the oracle, and electricity costs by up to 60\% relative to signal-agnostic baselines.

arXiv commentFull technical report for ACM SoCC 2026 paper with a small name change