Improvement For Kubernetes Scheduler

Modern data centers always run multiple different and complex applications. A powerful and stable resource scheduler is more and more significant to data center’s daily operation. The scheduler that manage all the resources in cloud and data center can be designed as the abstract layer between applications and cloud or data center independently. This is the so-called cloud scheduler or data center operating system. The emergence of container-as-a-service platform Kubemetes obviously reduce the difficulty of achieving this. However, both Kubemetes scheduler module’s architecture design and default algorithms are oversimplified which make it unable to meet users’ various demands and handle complex runtime environment.On the base of diving into Kubemetes overall architecture and its scheduler’s work flow, this article extends the system’s default algorithm library, puts forward to applying preemptive model to Kubemetes scheduler, reconfiguring jobs’resource quota and high availability resolution for centralized resource scheduler for the first time. Finally, I intergrate all designs into a single optimized solution to Kubemetes scheduler. Preemptive scheduling model breaks originally situation of single jobs’class and provides faster real-time response for some jobs such as web service.Reconfiguration for Jobs’resource quota reduces a waste of resource and improves security of resource supply. High availability resolution for centralized resource scheduler resolves single point failure of Kubemetes scheduler and improves system’s tolerance. Experiment data shows that my solutions to Kubemetes scheduler all bring significant availability and performance improvement.