CASPIAN: A Carbon-Optimized Multi-Cluster Job Scheduler

Abstract

It is no secret that AI/ML jobs utilize large number of power-hungry resources for extended periods of time, thus consuming exorbitant amount of energy. In this talk we will describe CASPIAN, an optimized carbon-aware multi-cluster job scheduler, which minimizes the carbon footprint of running jobs, without compromising their completion times. The CASPIAN scheduler runs in a control-plane cluster and uses open source projects such as MCAD and KubeStellar to dispatch and manage jobs in multiple workload clusters. Further, we will demonstrate through experimental results that CASPIAN does effectively reduce the carbon emission associated with computational energy consumption.