Adaptive Swapping for Variable Workloads in Real-time Task Scheduling

Energy conservation is one of the most important missions in the design of battery-based IoT systems. Recently, memory power consumption increases rapidly in real-time systems as the task size continues to grow and all tasks should reside in main memory to satisfy deadline constraints. Swapping is a well-known solution to save memory energy by flushing inactive tasks to storage and hibernating a certain part of memory. However, real-time systems do not allow swapping as swapping makes the prediction of execution time difficult. In this paper, we suggest a new swapping policy for real-time systems. To support swapping with real-time constraints, we adopt fast NVRAM storage and define a task model that characterizes swapping latency and energy precisely. We then locate inactive tasks temporarily in NVRAM instead of fully keeping all tasks in memory. As our policy optimizes the swapping conditions for all task sets in advance, it adapts to workload variations instantly and guarantees deadline requirements. Our simulations show that the proposed swapping policy reduces the memory energy consumption by 37.1% on average.