Speaker:Mohammad

Toward Real-Time Energy-Aware Automation of the Resource Scheduling Using Reinforcement Learning

Parallel real-time energy-aware scheduling is a challenging open-ended problem dealing with distributing parallel tasks represented as directed acyclic graphs (DAGs) among multiple processing cores while maintaining energy efficiency and minimizing response times. The problem has been addressed with non-learning and learning-based heuristics, but little consideration has been given to individual core configurations and temperature behavior when allocating tasks to cores. Additionally, the proposed learnable approaches did not address reward calculation, tuning, or model selection strategies suitable for sample-efficient high-dimensional multi-core environment action spaces. Consequently, this research proposes a practical online learnable framework for estimating and adjusting the per-core frequency configuration and temperature at runtime. Our method uses a hierarchical multi-agent cooperative reinforcement learning model in which one agent observes the profiler data to determine the frequency and the number of cores, and another agent gives priority to the cores based on average and differences in temperature observation. A non-policy dueling double deep Q-network is used to train these two agents to avoid over-fitting and overestimation issues while maintaining the sample-taking efficiency, and the network is made as shallow as possible to reduce the latency overhead. The proposed approach is compared with all the available Linux governors and a Federated energy-aware scheduler based on important parameters such as response time, energy consumption, average temperature, etc. An OpenMP DAG benchmark was used as the basis for these experiments using Intel Xeon 12-core and Intel Core i7 4-core processors. The results of real-time experiments on the Intel Core i7 show up to 26% improvement in energy consumption and 16% improvement in execution time compared to all other governors, and in the Intel Xeon processor, we have more than 3 degree C average temperature reduction compared to the state-of-the-art Linux governors.

Committee: Ali Jannesari (co-major professor), Abusayeed Saifullah (co-major professor), and Wensheng Zhang

Join on Zoom:

    Please click this URL to start or join. https://iastate.zoom.us/j/99732553933

    Or, go to https://iastate.zoom.us/join and enter meeting ID: 997 3255 3933