[Talk Proposal] SchedAI: AI-Assisted Linux Scheduling

19 views

Skip to first unread message

Aaryan P

unread,

Mar 27, 2026, 7:06:26 AM (5 days ago) Mar 27

to Kernel Meetup Bangalore

Hi,

I’d like to submit a talk proposal for the upcoming Bangalore Linux Kernel Meetup. I’d be happy to share some of the work I’ve been doing on Linux scheduling.

Title: SchedAI: AI-Assisted Linux Scheduling
Format: Lightning Talk (10 + 5 minutes)

Abstract:
Modern Linux systems increasingly need to balance performance and energy efficiency, especially on client devices where workloads are diverse and power consumption matters. Static scheduling choices such as fixed affinity or manual tuning often fall short because workload behaviour changes continuously over time. SchedAI explores a different approach: using machine learning to understand process behaviour and translate that insight into scheduler-friendly hints that work with the Linux kernel rather than against it.

SchedAI analyses kernel telemetry such as process scheduling statistics, CPU topology, and runtime activity patterns to classify workloads into categories such as light, moderate, and heavy. Based on this intensity estimation, tasks are guided towards the most suitable cores—performance cores for heavier workloads and energy-efficient cores for lighter ones—while minimizing unnecessary migrations and improving overall CPU utilisation. The system applies Linux-native controls such as UCLAMP and nice values to influence scheduling decisions in a lightweight and adaptive manner. The result is a practical framework that improves performance, as demonstrated through benchmarking, while also accounting for actual energy consumption.

In this talk, I will walk through the SchedAI architecture, the data pipeline behind the model, the hinting strategy used to influence scheduling, and the lessons learned while integrating AI techniques into Linux kernel workflows. The goal is to show how ML can complement the scheduler by providing useful runtime guidance without requiring invasive changes to kernel behaviour.

Outline:

Motivation: Limitations of static scheduling and the need for adaptive, workload-aware decisions
Overview of SchedAI architecture and design goals
Kernel telemetry and feature extraction for workload understanding
ML-based workload classification (light, moderate, heavy)
Mapping workload intensity to scheduling decisions across heterogeneous cores
Translating predictions into scheduler hints
Demo: running SchedAI on real workloads and observing scheduling behaviour
Benchmarking results: evaluating performance improvements and energy impact using standard tools
Practical insights: what worked, trade-offs observed, and future directions

Speaker Bio:
I am a Platform Software intern at AMD with a strong interest in Linux kernel internals, particularly scheduling and performance optimization. My work focuses on applying machine learning to real-world systems problems. I collaborated with Vishal Badole, MTS, Software System Design Eng., in designing and building SchedAI. This work reflects my broader interest in combining systems engineering with AI to develop practical solutions that improve performance, efficiency, and resource utilization in modern operating systems.