Kinitos: Dynamic network-aware scheduling and descheduling in Kubernetes clusters with mobile nodes

The current Cloud/Edge/Fog platforms, available in practice, do not support dynamic management of computational and network resources in clusters with moveable and constraint nodes, like in applications such as: connected vehicles, Internet of Things (IoT), spacecraft computing, etc. This leads to increased total response time and low QoS for the long-living applications composed of many inter-dependent microservices with stringent and configurable network latency and bandwidth requirements. This problem is further complicated by the mobile end-users. The paper presents a novel framework for network-aware dynamic scheduling and descheduling (migration) of containers on mobile clusters based on Kubernetes (K8s), named Kinitos. It is composed of plugins in the K8s scheduler, descheduler and Application Group and Network Topology controllers, providing a mechanism for configuration of application microservice dependencies with specific computational (CPU, memory, storage) and network (latency, bandwidth) requirements. In clusters with dynamic, moveable and constraint nodes, it migrates application containers on optimal nodes to satisfy network and computational requirements continuously at runtime. The performed experiments with real world node movement patterns on K8s ARM64 testbed cluster show an improvement in the total end-to-end network latency of a long-living edge-native application by up to 60% compared to the K8s default scheduler and up to 50% compared to the latest K8s network-aware scheduler plugin, with only 0.55 s (32.5%) increase in the average pod deployment time. Its implementation can be used not only for theoretical research, but in practice.