Real-time Scheduling in Datacentre Clusters

Sammanfattning: Industry 4.0 can be described as the next generation-factories that is characterised by putting a high demand for automation and flexible production lines. The proposed way to achieve this goal is through a large number of Industrial IoT devices(IIoT) in the factory, some having high availability- and low-latency requirements. This demands for a software that can monitor and manage the applications used by these devices. The current industry standard for managing clusters of applications is called Kubernetes. However, using Kubernetes in this environment means that it needs to meet the low latency demands of Industry 4.0. The purpose of this thesis was to investigate the possibility of achieving this support in Kubernetes. To investigate if this was possible, a Kubernetes cluster was deployed to Ericssons private cloud Xerces. On this cluster services was deployed that executed arbitrary tasks with the invocation of an HTTP-request. The difference in how fast these tasks executed was used as a metric to see the delay some request were experiencing due to non optimal scheduling on Kubernetes. Through investigations of how the underlying kernels schedules jobs on it’s CPU proposed solutions to reduce the execution time was made. These solutions were then tested on the cluster and compared with the first measurements. The results of these measurements showed a reduction in execution time for the deployed services with the introduction of the preemptive scheduling policy SCHED_FIFO in Linux, which uses the first in first out algorithm. This improvement in execution time was at the cost of a degradation of the response time of the cluster. These findings points towards the conclusion that applying a real time scheduling policy to processes on a Kubernetes cluster is not cost free. However, a suggested path for how to further optimise a Kubernetes cluster in regards to response time has been proposed.