A flattened hierarchical scheduler for real-time virtualization

Migrating legacy real-time software stacks to newer hardware platforms can be achieved with virtualization which allows several software stacks to run on a single machine. Existing solutions guarantee that deadlines of virtualized real-time systems are met but can only accommodate a reduced number of systems. Therefore, this paper introduces ExVM, a new scheduling framework to maximize the number of legacy uniprocessor real-time systems able to run on a single machine. Contrary to most existing solutions, ExVM uses a flattening approach where the host schedules the virtual machine which contains the task with the earliest deadline. The real-time characteristics of tasks are obtained through introspection during the execution. We implemented this framework using Linux's SCHED DEADLINE real-time scheduling policy in the host. Simulations using an exact schedulability test show that ExVM is able to schedule 96% of randomly generated tasksets with a utilization of at least 0.8, while state-of-the-art solutions are only able to schedule 40% of the same tasksets. Experimental evaluations performed using synthetic benchmarks and production real-time applications show that ExVM always outperforms the existing solutions, always meeting more than 80% of deadlines while these solutions fall below 50% when the utilization increases.