Anastacia Grishchenko: Synchronization Strategy for Hardware Tasks to Improve Energy Efficiency (Großer Beleg)

26.04.2023, 14:00 pm

Invitation to the presentation of Ms. Anastacia Grishchenko

Topic: Synchronization Strategy for Hardware Tasks to Improve Energy Efficiency

Project: Großer Beleg

Supervisor: Cornelia Wulf

Abstract: With the broader range of applications in which Field-Programmable Gate Arrays (FPGAs) can be used and the growing demand for low-power applications, there is clearly a stronger urge to reduce the power dissipation of FPGAs, which are traditionally not as energy-efficient as Application-Specific Integrated Circuits (ASICs). At the same time, in recent years there is a tendency to run multiple applications or even operating systems on a single FPGA, which requires sharing the resources available on the platform. One of the common methods to reduce the power consumption of FPGAs is Dynamic Voltage Scaling (DVS), which allows dynamically modifying the voltage, e.g., to a level just sufficient to ensure correct operation. However, DVS can only be applied to the entire platform and not to individual resources, such as the hardware accelerators. Consequently, some hardware accelerators might run at a higher voltage than is actually required to perform the assigned task, resulting in unnecessary power consumption. To tackle the presented problem, this work proposes a synchronization and scheduling algorithm. The algorithm modifies the schedule of the hardware tasks taking into account their real-time constraints and the minimum voltage required to operate correctly. The main goal
is to execute the tasks as parallel to each other as possible, so that each task with a lower
voltage requirement can eventually benefit from the higher voltage required by other tasks. Moreover, the algorithm was integrated into the real-time operating system FreeRTOS, which was responsible for scheduling the allocated hardware tasks on the accelerators. For a task set with sufficient intersection of execution times between tasks and not overly restrictive deadlines, improvements up to 5.26% and 3.24% were achieved when compared to the scenario where not optimizations were performed at all and to the scenario were the tasks were not synchronized before applying DVS.