Load Balancing Prioritized Tasks via Work-Stealing

Work-stealing schedulers focus on minimizing overhead in task scheduling. Consequently, they avoid features, such as task priorities, which can add overhead to the implementation. Thus in such schedulers, low priority tasks may be scheduled earlier, delaying the execution of higher priority tasks and possibly increasing overall execution time. In this paper, we develop a decentralized work-stealing scheduler that dynamically schedules fixed-priority tasks in a non-preemptive manner. We adhere, as closely as possible, to the priority order while scheduling tasks by accepting some overhead to preserve order. Our approach uses non-blocking operations, is workload independent, and we achieve performance even in the presence of fine-grained tasks. Experimental results show that the Java implementation of our scheduler performs favorably compared to other schedulers (priority and non-priority) available in the Java standard library.