쿠버네티스 환경에서 이벤트 기반 오토스케일링 성능 분석

초록

This paper quantitatively compares and analyzes the performance characteristics of the Kubernetes Event-driven Autoscaler (KEDA) under various workload conditions to propose performance comparison criteria for establishing efficient autoscaling strategies in Kubernetes environments. Four KEDA-compatible message brokers - RabbitMQ, Apache Kafka, Redis Streams, and NATS JetStream - are evaluated through autoscaling experiments for three workload types: burst workload, sustained workload, and decreasing workload. The experiments are conducted in a controlled Kubernetes environment, where TCP-based workload generators are used to measure autoscaling behavior. The analysis focuses on two major dimensions: reaction characteristics and resource utilization. Based on these experimental results, this paper presents the autoscaling performance characteristics of KEDA with different message brokers under varying workload conditions.

키워드