GreenCloud: Carbon-Aware Resource Scheduling in Cloud Data Centers using Deep Reinforcement Learning

Cloud data centers are large power consumers
and thus substantial contributors to carbon emissions.
Conventional resource scheduling techniques mostly focus on
performance and energy efficiency, mostly ignoring the
environmental footprint. This work introduces GreenCloud, a
new carbon-conscious resource scheduling system using Deep
Reinforcement Learning (DRL) to reduce carbon emissions in
cloud computing systems. By integrating real-time carbon
intensity information into decision-making, GreenCloud
adaptively distributes workloads among geographically
dispersed data centers both according to resource and
environmental conditions. Empowered with Proximal Policy
Optimization (PPO) as the backbone of the DRL algorithm,
our model learns efficient scheduling policies that optimize
energy usage, carbon emissions, and SLA compliance.
Experimental results show GreenCloud to have the ability to
curtail carbon footprints by as high as 56% lower than Round
Robin scheduling yet retain high utilization of resources and
low SLA violation rates.