A real-time distributed resource-granting task scheduler for a virtual infrastructure management system

Abstract

This M.Eng. thesis introduces a distributed real-time task scheduler that is designed and implemented to replace the role of a lock server for a cloud service that manages data centers that contain virtual machines. A round-robin scheduling algorithm is proposed that orders tasks based on their needs of locks, serializing tasks that conflict in their lock use. Tasks are required to obtain all locks before execution. The distributed part of the scheduler is based on Zookeeper, an open-source data service for distributed applications. The scheduler includes a failover scheme that relies on task states stored on Zookeeper. The design is highly available-the scheduler will keep scheduling tasks as long as at least one instance is running. Experimental results show that the scheduler cuts down overall task completion time by approximately a half compared to a lock server on both benchmarks used.