Energy-efficient performance optimization in Kubernetes microservices using Generalized Stochastic Petri Net

The advantages of microservices system architectures orchestrated by Kubernetes include their capacity to adjust dynamically to meet varying demand, thereby ensuring application performance requirements during high load periods while reducing electrical consumption during low demand periods. However, the proper configuration of autoscaling in microservices systems presents significant challenges due to the myriad of parameters involved, with optimal choices highly dependent on both the application and the underlying infrastructure. Balancing system performance with energy efficiency is inherently difficult, as enhancements in one area often lead to detriments in the other. In this article, a model is presented to aid in the planning of performance and electrical consumption for microservices architectures orchestrated by Kubernetes, utilizing Pod and cluster autoscaling. The approach is based on the application of Generalized Stochastic Petri Net models. This methodology enabled the creation of a comprehensive model that incorporates various Kubernetes autoscaling configuration elements, the performance characteristics of individual microservices, and the capacity of the underlying infrastructure. The proposed model is capable of computing key metrics, including response time, throughput, discard probability, electrical consumption, average consumption per request, and the Energy-Response time Weighted Product. Sensitivity analysis was employed to identify the elements exerting the greatest impact on the performance and electrical consumption of the architecture, thereby guiding parameter adjustment efforts. It was also determined that the optimal configuration choices are contingent on the expected workload. The results indicate that using configurations with higher autoscaling thresholds during low workloads can reduce electrical consumption by approximately 32% without significantly degrading performance. Conversely, in high arrival rate scenarios, this autoscaling configuration results in a 37% reduction in consumption but at the cost of a 175% increase in response time.