VILNIUS TECH Library invites you to follow the published new dissertations. The dissertation „Research of service level agreement aware autoscaling algorithms for containerized cloud-native applications“ prepared by VILNIUS TECH, Olesia Pozdniakova. The dissertation was prepared in 2016–2025. Scientific Consultant – Prof. Dr Dalius Mažeika.
The dissertation was defended at the public meeting of the Dissertation Defense Council of the Scientific Field of Informatics Engineering in the Aula Dotoralis Meeting Hall of Vilnius Gediminas Technical University at 1 p.m. on 16 May 2025.
The development of cloud-native applications focuses on scalability and loose coupling of containerized microservices to ensure smooth deployment on cloud or container orchestration platforms. An autoscaler is a crucial component responsible for dynamically provisioning compute resources. When dynamically provisioning resources, addressing issues such as timelines and the amount of resources to be provisioned is important. Therefore, most autoscaling algorithms aim to find a balance between avoiding Service Level Agreement (SLA) violations and effectively managing costs or energy. Various rules-based autoscaling approaches were created to address quality of service concerns and minimise the risk of SLA violations. When resources are allocated and adjusted as needed, an autoscaler typically evaluates current service performance by comparing it to a predefined service level indicator (SLI) value. However, this alone may be insufficient to address changes in SLA conformance. To respond appropriately, the autoscaler must also consider the system’s overall SLA fulfillment status. This research presents two innovative self-adaptive autoscaling solutions for SLA-sensitive applications. The first solution focuses on maintaining the defined Service Level Objective (SLO) to recover from service degradation and achieve the desired service level. The second solution features a novel SLA-aware dynamic CPU threshold adjustment algorithm. The algorithm aims to ensure that the application has sufficient resources to operate at a level that keeps the number of response time violations compliant with the SLO. Additionally, it aims to ensure that the system operates as closely as possible to the defined Service Level Objectives, thus minimising resource wastage. The solution employs exploratory data analysis techniques in conjunction with moving average smoothing to determine the target utilisation threshold. The Kubernetes Horizontal Pod Autoscaler (HPA) remains the most widely used threshold-based autoscaling due to its simple setup, operation, and seamless integration with other Kubernetes functionalities. For that reason, this research compares the autoscaling solutions proposed here with the Kubernetes Horizontal Pod Autoscaler and evaluates their effectiveness and performance across various real-world workload scenarios. The evaluation methods for algorithms focus on their ability to operate near-defined SLOs and the effectiveness of resource provisioning. The analysis of the experimental results demonstrates that these solutions are successful in SLA fulfillment and SLO restoration goals while providing an adequate amount of resources to achieve these objectives. The results of the dissertation were published in six scientific publications, two of which were in reviewed scientific journals indexed in Web of Science and presented at five international conferences.
Doctoral dissertation readers can search via VILNIUS TECH Virtual Library.
