KubeDSM is a Kubernetes scheduler module designed to operate on cloud-assisted edge clusters. The following figure displays an overview of the cluster:
You can find more details about this scheduler and the experiments in our paper. You can access the paper here: KubeDSM.
KubeDSM's goal is to minimize user response time by using edge resources as much as possible. To do so, KubeDSM is dynamically migrating pods from cloud to edge. Also, it will migrate pods between edge nodes, to reduce resource fragmentation. This will help the upcoming pods to fit better in the edge. The scheduler is also able to do batch scheduling (i.e. scheduling multiple pods simultaneously). You can see the scheduler's component diagram in the figure below:
The underlying scheduling algorithm takes advantage of the edge cluster's small resource size to make almost optimal decisions to maximize its goal function (Quality Of Service, QoS). The algorithm's data flow diagram:
The scheduler's performance has been experimented with various generated multi-pattern workloads for hours in a heterogeneous edge cluster connected to a cloud node. The results show that KubeDSM achieved a 50% reduction in user response time, a 20% increase in resource utilization, and fulfilled 100% of theoretically satisfiable QoS constraints.
The following is the list of the projects developed alongside KubeDSM for this research:
- Genny: Genny generates scenarios based on the given config. It is used to generate normal distributed and wavy scenarios.
- DrStress: DrStress implements the given scenario to generate workload for the scheduler.
- MineDraft: MineDraft extracts draft diagrams and tables from the metrics aggregated during the scenario execution, these drafts are further used to analyze and create diagrams for the paper.
- Sencillo: As Kubernetes default scheduler does not perform well in heterogeneous edge clusters, several scheduling algorithms are implemented in Sencillo to compete with KubeDSM in the experiments.