engleski

Scalability analyses of number and metric types for monitoring Docker container systems managed by Kubernetes platform

Kubernetes is a framework used to orchestrate Docker containers. Due to its applicability in using with microservices and container virtualisation, Kubernetes continues to gain popularity among medium and large software development teams. Microservices architecture have inspired Kubernetes to promote developing software systems in small, independent executable parts called Pods, thus enabling quick component replacement and scalability. In the microservice architectures, dominating approach to virtualisation is based on using containers, and as such offers quick and standardised deployment of Kubernetes system. This master thesis focuses on exploring the scaling capabilities of the Kubernetes system utilizing Horizontal Pod Autoscaler (HPA) as one of the key Kubernetes components. Firstly, the main concepts of Docker concept, Kubernetes framework, Grafana and Prometheus tools are explained. Kubernetes platform has been presented in detail in terms of all Kubernetes components including their interactions and implementation, with demonstrations based on a local cluster composed of multiple Nodes. The concurrent software tools of the Kubernetes platform are also explained with some necessary practices implemented with the tested Kubernetes platform. In order to adapt the system to different working loads, Kubernetes offers the ability to change the number of Pod replicas through concept of resource auto-scaling. This is done using HPA which performs auto-scaling of resources based on the average Pod CPU and memory metrics obtained from the metrics-server. Auto-scaling was done on a web server deployed as a PHP-apache Pod with an index page containing complex CPU tasks. A traffic generator is deployed as a “curl” command inside a Pod and configured to display request-response time, which is used as an additional metric to display responsiveness of a system concerning the number of PHP-apache replicas. The autoscaling experiments were ran using different parameters, such as single and multi-node clusters, with a variable number of Pod replicas. The data obtained from running the experiments are collected with Prometheus and displayed using Grafana. Afterward, an analysis is made using the collected data with the results showing that increasing the number of PHP-apache replicas leads to reduced request-response time and reduced average CPU load across all PHP-apache Pods. Additional observations are made confirming that there exists a waiting period after a replica Pods scale-up operation and before a scale-down operation, which is necessary for ensuring system stability. Also, results show that the auto-scaling of Pods number has an impact on the instantaneous temperature of CPU, with temperature decreases as the number of replicated Pods increases and vice versa. Lastly, the conclusion confirms Kubernetes effectiveness in scaling the deployment of a system based on microservices architecture with containers and emphasizes the importance of the Kubernetes framework in today's software development.

Docker ; Kubernetes ; Prometheus ; Grafana ; Horizontal Pod Autoscaler ; Metrics ; Replicas ; PHP-Apache ; Services ; Deployments

nije evidentirano