P3: A task migration policy for optimal resource utilization and energy consumption

Abstract

Theevolution in modern technologies like artificial intelligence, machine learning, cloud computing, edge computing, data science, etc, focuses on user perspectives like accuracy, response-time, and timeliness but at the same time consumes heavy energy due to large and fast data processing. From the system perspective, resource utilization and energy consumption are also significant design considerations. This work proposes a task migration policy for optimal core utilization and energy savings. The time taken by data analytical tasks to process the data varies, due to variations in the amount of data it analyzes in unit time. This creates variation in the core utilization due to which there exist small inactive intervals in the schedule, consuming energy. If the inactive state is known to continue for a longer duration, the core can be put into a shutdown state which effectively reduces overall energy consumption. Dynamic Procrastination (DP) is a commonly used technique to increase the inactive duration by postponing the tasks whenever possible. To further increase the inactive duration to qualify for shutting down the core, in a homogeneous multi-core (HMC) system, the jobs can be migrated to other cores. This effectively improves core utilization and reduces overall system energy without negatively affecting performance. Combining the DP and migration techniques introduces challenges like meeting deadlines, deciding upon push/pull migration, finding the number of tasks and suitable core for migration, and computation of energy consumption parameters. This paper proposes P3 (Push-Procrastinate-Pull) migration policy for the HMC system. The experimental evaluation with synthetically generated benchmark program suites shows that on an average P3reduces the overall energy by 1.2% and reduces the shutdown duration over the idle period by 2.22% over DP without migration.