Oleg I. Gorbatov

The serverless computing paradigm has transformed cloud infrastructure by favoring dynamic resource allocation and automatic scaling, reducing the complexity of infrastructure management. However, as workloads with high computing demands and rapid instantiation requirements become more prevalent, energy efficiency remains a critical challenge. Inefficient workload scheduling can lead to suboptimal resource utilization and increased power consumption. In this respect, traditional scheduling policies often struggle to adapt to the dynamic and unpredictable nature of serverless workloads, highlighting the need for more energy-aware scheduling strategies. In this work, we propose a Reinforcement Learning (RL)-assisted scheduling approach to enhance energy efficiency in serverless cloud computing. We develop a Deep Q-Network (DQN)-based scheduler that continuously learns workload placement strategies, minimizing cluster-wide power consumption, and consolidating workload across available Nodes. We implement our approach as a custom scheduling plugin for Kubernetes, ensuring seamless integration with Knative-based serverless workloads. We evaluate our approach using live system telemetry and compare its performance against baseline scheduling techniques. The results show that our RL-based scheduler outperforms the default Kubernetes scheduler by 9.5 % in the total cluster CPU consumption, allowing potential energy savings.

Energy efficiency is becoming increasingly important for sustainability. In a cloud-native environment, orchestration platforms such as Kubernetes can significantly impact the overall energy performance of applications as they manage their lifecycle, including scheduling. Kubernetes, specifically as a de facto cloud platform, has a default scheduler that does not account for energy consumption in its decision-making process, highlighting the need to explore mechanisms that can improve the energy performance of deployed applications. In this paper, we present an energy-aware scheduling mechanism within Kubernetes to optimize the energy usage of cloud-native applications while meeting other scheduling requirements and constraints. This approach extends the scheduling logic in Kubernetes through plugins that integrate energy metrics and employ Machine Learning-based power estimation for energy-optimized decision-making. The evaluation of the proposed approach shows up to 12 % energy savings with minimal performance degradation. When complemented by other consolidation techniques, it could potentially lead to substantial savings ranging from 21 % to 65 %, depending on the system load in the scenario analyzed. Importantly, this method does not introduce significant overhead.