Dynamic Scheduling of Shared Resources using Reinforcement Learning

Abstract: The goal of the thesis is to simulate the Ericsson Many-Core Architecture, EMCA, and implement a dynamic scheduler for the system using reinforcement learning methods. The system contains shared resources that receive and complete jobs. Also, the deadlines and latency definitions can change depending on the job type. The scheduler should aim to avoid missing deadlines as well as aim to reduce the overall latency in the system. A python simulation has been implemented of the Ericsson Many-Core Architecture and two reinforcement learning based schedulers have then been developed and used for different configurations. They are evaluated by comparing their performance to a random and a static scheduler. The first scheduler uses Q-learning and the second uses a version of Q-learning with a neural net that approximates the Q-function. The results showed that the second version experienced issues with convergence which caused deadline misses and poor latency. The regular version of Q-learning showed promising results, avoiding deadline misses and was able to reduce the latency below that of the static scheduler for one of the systems. There are still some issues that could be addressed as well as avenues to explore regarding the scheduler. Furthermore, in order to apply the scheduler to the real system, some modifications are necessary. However, the simulations show that the reinforcement learning can successfully be used as a scheduler on the EMCA for different configurations.