Date of Award

6-10-2026

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2026

Degree Name

Master of Science (MS)

Department

Computer Science and Engineering

First Advisor

Yidi Wang

Abstract

Modern Linux systems ship with a single default CPU scheduler that remains active regardless of how workloads change over time. Although the sched ext framework permits loading alternative scheduling policies at runtime, existing approaches to selecting among them rely on manual configuration rather than automatic adaptation to workload characteristics. We present an automated scheduler selection system that monitors workload behavior and dynamically switches between scheduling policies. Using discrete-event simulations of four Linux scheduling policies, we formulate scheduler selection as a sequential decision problem with a composite reward signal covering latency, fairness, throughput, and CPU utilization, and apply a two-phase behavioral cloning approach: reward-driven discovery identifies a preferred scheduler for each workload type without human labels, and a neural network is then trained via supervised learning to map runtime observations to the discovered actions. In our evaluation on simulated workload scenarios, the resulting policy selects the expected scheduler in each workload phase and generalizes to irregular phase durations without retraining.

Download

Share

COinS