Smart scheduling : optimizing Tilera's process scheduling via reinforcement learning

Abstract

As multicore processors become more prevalent, system complexities are increasing. It is no longer practical for an average programmer to balance all of the system constraints to ensure that the system will always perform optimally. One apparent solution to managing these resources efficiently is to design a self-aware system that utilizes machine learning to optimally manage its own resources and tune its own parameters. Tilera is a multicore processor architecture designed to highly scalable. The aim of the proposed project is to use reinforcement learning to develop a reward function that will enable the Tilera's scheduler to tune its own parameters. By enabling the parameters to come from the system's "reward function," we aim eliminate the burden on the programmer to produce these parameters. Our contribution to this aim is a library of reinforcement learning functions, borrowed from Sutton and Barto (1998) [35], and a lightweight benchmark, capable of modifying processor affinities. When combined, these two tools should provide a sound basis for Tilera's scheduler to tune its own parameters. Furthermore, this thesis describes how this combination may effectively be done and explores several manually tuned processor affinities. The results of this exploration demonstrates the necessity of an autonomously-tuned scheduler.