Modelling, Characterising & Scheduling Services in Kubernetes

What is it about?

This work presents a study on modeling performance and resource management within the Kubernetes (K8s) system, recognizing the role of containers versus virtual machines in cloud-based implementations. The authors propose a formal analytical model based on Reference Nets (a type of Petri Net) to simulate the lifecycle and operational states of K8s pods and containers. This model is designed for developers to evaluate application performance and capacity planning. To characterize the model, experiments are conducted measuring critical variables such as container creation and termination times, extracting empirical data from a real K8s implementation. Additionally, micro-benchmarks are performed comparing the impact of different configurations for CPU, I/O, and network-intensive applications. The benchmark results indicate that CPU-intensive applications benefit from grouping all containers into a single pod, while network performance is improved by deploying multiple isolated pods.

Featured Image

Photo by Growtika on Unsplash

Photo by Growtika on Unsplash

Why is it important?

The development of this formal model, characterized using real data from a Kubernetes implementation, generates a significant impact that could increase its readership by offering practical tools: 1. Design of Scalable Applications: The Reference Net model can be used as a basis for designing scalable applications that utilize Kubernetes. 2. Performance Evaluation: Developers and application designers can use the model to assess how pods and containers might impact their application's performance. 3. Capacity Planning: The model can be employed to support capacity planning and resource management for application scaling. This formal approach and its practical results are crucial for a system like AWS Lambda (a cited example of serverless computing), where resource provisioning occurs at 100 ms intervals. Understanding the performance associated with the deployment and termination of containers is fundamental for a provider to offer more granular billing options to users with streaming analytics requirements. The uniqueness and timely relevance of this work lie in its analytical approach and direct application to Kubernetes in the context of the rapid adoption of cloud containers. Unique and Timely Aspects of the Research This work is timely because containers are rapidly replacing Virtual Machines (VMs) as the preferred computing instance in cloud-based implementations, mainly due to the significantly lower operational costs associated with their deployment. This trend is driven by interest in cloud-native applications (CNA) that require support for scalability and resilience, as well as emerging classes of applications such as streaming processing that need to launch and terminate instances on a "per second" basis. For such applications, the overhead associated with VMs remains a limitation. What distinguishes this research is its rigorous analytical approach to addressing performance issues with containers. Specifically, the authors develop a performance and management model based on Reference Net (a type of Petri Net) for Kubernetes. The authors note that, although the virtualization of VMs and containers has attracted considerable attention regarding performance comparison, to their knowledge, the performance of containers has not been addressed using a rigorous analytical approach or formal models. Difference That Could Make a Mark and Impact on the Audience The development of this formal model, characterized using real data from a Kubernetes implementation, generates a significant impact that could increase its readership by offering practical tools: Design of Scalable Applications: The Reference Net model can be used as a basis for designing scalable applications that utilize Kubernetes. Performance Evaluation: Developers and application designers can use the model to assess how pods and containers might impact their application's performance. Capacity Planning: The model can be employed to support capacity planning and resource management for application scaling. This formal approach and its practical results are crucial for a system like AWS Lambda (a cited example of serverless computing), where resource provisioning occurs at 100 ms intervals. Understanding the performance associated with the deployment and termination of containers is fundamental for a provider to offer more granular billing options to users with streaming analytics requirements. In essence, this work provides the missing bridge between emerging container technology (Kubernetes) and rigorous formal methods (Petri Nets), offering concrete tools for optimization and cloud system design.

Perspectives