Adapting Containerized Workloads for the Continuum Computing

Container and microservices management platforms are currently one of the most important tools for cloud computing, but since the scope of these tools is homogeneous cloud architectures they have serious limitations in adapting to new computing paradigms. Therefore, using the default scheduler in heterogeneous node systems faces significant limitations when tasked with orchestrating workloads in a Continuum Computing environment, as the nodes have very different characteristics and restrictions. To solve this limitation we decided to use Kubernetes as it is the most popular Container management tool and we propose to replace the native scheduler with a reimplementation that gives us complete flexibility for the process of assigning pods to nodes, providing a framework to design algorithms that considers all the necessary parameters for the deployment of services in a Continuum. In addition, we address one of the most limiting aspects of the K8s scheduler, its pod-by-pod allocation approach, which makes it difficult to optimise the complete set of allocations. To test our proposal we design a use case and perform several tests on a real environment based on virtual machines, in which stress tests are conducted to measure the performance of each method. We then present a series of results to justify the benefits of our proposal, including the reduced performance provided by the pod-pod approach and how a batch-based approach greatly improves efficiency. The results show the usefulness of using batch-based approaches and how the Kubernetes scheduler extension points are not enough to support the requirements of the Continuum.