Date of Award
Doctor of Philosophy (PhD)
Modern real-time applications are becoming more demanding computationally while their temporal requirements, dictated by the physical world, often remain unchanged. This coupled with the increasing prevalence of multiprocessors in real-time systems necessitates that highly computation-demanding real-time tasks need to be parallelized to exploit the parallelism offered by the underlying hardware, in order to satisfy their temporal constraints. Scheduling parallel real-time tasks, however, introduces a new layer of complexity due to the allowance for intra-task parallelism. This dissertation addresses the problem of scheduling parallel real-time tasks in which tasks may (or may not) access shared non-processor resources, such as in-memory buffers or data structures. Specifically, for independent tasks, we propose new scheduling algorithms and schedulability analyses for parallel tasks with these characteristics, under federated and global scheduling.Experimental results show that the proposed algorithms and analyses improve the previously introduced methods. For parallel tasks that may access shared non-processor resources, we present a blocking analysis for two different types of spinlocks; through evaluations, we make a recommendation for a preferable ordering of locks. We also study practical runtime parallel scheduler designs for soft real-time applications and present a design that is more suitable for soft real-time systems.
Christopher D. Gill
Christopher D. Gill, Kunal Agrawal, Sanjoy Baruah, Roger Chamberlain,