Reinforcement Learning based Fragment-Aware Scheduling for High Utilization HPC Platforms

Abstract

Due to high capacity and complex scheduling activities, a HPC platform often creates resource fragments with low usability. This paper develops a novel fragment-aware scheduling approach which improves system utilization by fitting elastic lightweight tasks to the fragments of resources dynamically. The new approach employs a threshold to determine the balancing factor between the length of tasks and the degree of granularity of the resource fragments. We employ the PPO reinforcement learning approach to train a neural network that can compute the threshold precisely. With the threshold that is adaptive to the changing system states, the PPO-based scheduler is able to utilize the idle resources and maximize the execution success rate of the tasks.