University of Illinois Urbana-Champaign

A machine learning approach for load balancing in the Linux kernel

Chen, Jingde

This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.

Permalink

https://hdl.handle.net/2142/107283

Description

Title
A machine learning approach for load balancing in the Linux kernel
Author(s)
Chen, Jingde
Contributor(s)
Iyer, Ravishankar
Issue Date
2020-05
Keyword(s)
  • Linux kernel
  • Completely fair scheduler
  • Operating system
  • Load balancing
  • Machine learning
Abstract
The default scheduler in the Linux Kernel - Completely Fair Scheduler (CFS) aims to provide fairness and maximized utilization of the processing time. It performs load balancing between cores by looking at the cache locality and CPU utilization of individual processes. However, memory cache and processing time are not the only critical hardware resources shared by processes. Contention due to imbalanced needs for other shared computing resources including memory bandwidth, IO bandwidth, arithmetic operations, etc. may still hinder the overall performance. Traditional logical approaches bear limitations in solving this problem due to complex relationships between resources and performance, especially when resources are shared on different levels on the NUMA and chip multiprocessor (CMP) architectures. This thesis explores the possibility of applying Machine Learning (ML) to implement resource-aware load balancing in the Linux kernel. We propose an approach of applying machine learning to a part of the Linux CFS load balancer by imitation learning. Dynamic tracing of kernel functions with eBPF and kprobes was used for runtime data collection of the system. A Multi-Layer Perceptron (MLP) model was then trained to emulate existing kernel load balancing decisions. We present both floating-point and fixed-point implementations of the model inside the kernel and evaluate their performance. Integration of processors’ Hardware Performance Counters (HPCs) for usages inside the kernel via the Perf Subsystem is studied as a support for future works in implementing resource-aware load balancing.
Type of Resource
text
Language
en
Permalink
http://hdl.handle.net/2142/107283

Owning Collections

Senior Theses - Electrical and Computer Engineering PRIMARY
The best of ECE undergraduate research