CTA: A Critical Task Aware Scheduling Mechanism for Dataflow Architecture

Critical tasks directly affect the overall performance of a program, especially in dataflow architecture. The reason is that dependency between tasks in dataflow scenarios is much more complex than those in control-flow scenarios. However, previous works fail to sufficiently accelerate the critical task because most of them only applied optimization for static critical tasks, without considering the runtime status during execution. We propose a critical task aware (CTA) scheduling mechanism for dataflow architecture. By adopting co-optimization of hardware and software, higher execution priorities are assigned to the critical tasks for better scheduling. The experimental results show that our mechanism increases the computational performance by 14%–78%, and increases the power efficiency by 11%–41%.