Caspian: A Carbon-aware Workload Scheduler in Multi-Cluster Kubernetes Environments

Abstract

The surge in demand for computing resources in data centers coupled with the rise of environmental concerns has motivated cloud providers to reduce carbon emissions due to computational energy consumption. An opportunity lies in the fluctuating availability of renewable energy over time and the variability of power sources over grid regions, leading to variations in space and time in carbon intensity. Exploiting such variations, this paper introduces Caspian, a carbon-aware workload scheduler in multi-cluster Kubernetes environments, which aims at reducing carbon emission associated with the execution of workloads, while satisfying Quality of Service (QoS) requirements. Caspian cooperates with the multi-cluster management platform to apply scheduling and placement decisions over a set of distributed clusters. We present efficient optimization algorithms to achieve these goals. Further, we describe an implementation of Caspian, integrated with Multi Cluster App Dispatcher (MCAD), a multi-cluster management platform which handles queuing and dispatching of workloads over multiple clusters. Our experimental results show that Caspian effectively reduces the carbon emission associated with the computational energy consumption, compared to a baseline scheduler which only satisfies the QoS of workloads. Specifically, Caspian could reduce carbon emission by about 33%, with about 98% of workloads completing at an average fraction of 0.6 of their deadline.