Energy-Aware Latency Optimization for Scheduling Serverless Workload in Edge Computing Environment

In order to accommodate latency-sensitive IoT and AI workloads, serverless computing is becoming more popular in edge environment. However, the default Kubernetes scheduler ignores the energy and performance limitations of edge nodes and is resource-agnostic. Prior approaches usually only optimized for latency or energy, ignoring the combined effects of cold-start dynamics, inter-node communication, and inter-service dependencies. In this work, we propose a lightweight heuristic scheduling approach that combines inter-service traffic, energy, and latency into a single cost function. This approach, implemented as a custom Kubernetes Scheduling Framework plugin, has low overhead and is used in conjunction with a descheduler that consolidates workloads by draining underutilized nodes. Short-term responsive placements and long-term energy efficiency are made possible by this combination. We test the system on a Raspberry Pi cluster, using Knative workloads that are typical of IoT analytics workflows. The average latency decreased by 29%, failure rates decreased by 74%, and energy consumption per request reduced by 32%, all of which are consistent improvements over the default scheduler. These results show that multi-objective, metrics-aware placement can significantly improve serverless edge platforms’ quality of service objectives and energy efficiency, specifically when combined with descheduling for consolidation.