Learning-based scheduling for ray-based Hybrid HPC-Cloud Systems

Lu, Yicheng

Permalink

https://hdl.handle.net/2142/124562 Copy

Description

Title

Learning-based scheduling for ray-based Hybrid HPC-Cloud Systems

Author(s)

Lu, Yicheng

Issue Date

2024-04-30

Director of Research (if dissertation) or Advisor (if thesis)

Kindratenko, Volodymyr

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Cloud Bursting
• Hpc
• Data Movement
• Scheduling

Language

eng

Abstract

Hybrid HPC-Cloud systems are becoming increasingly popular within the scientific community for their ability to efficiently manage sudden increases in demand, thus improving the processing times of HPC workloads. However, current systems lack efficient workload scheduling strategies to suit these hybrid environments and face considerable deployment challenges due to intricate configurations required, particularly concerning data transfer between HPC and the cloud. To address these issues, we have developed an innovative HPC-Cloud bursting system using Ray, a well-known open-source distributed framework. Our system adopts a learning-based scheduling approach at the function level through a dynamic label-based architecture and automatically manages data movement between the cloud and HPC. Specifically, our scheduler proactively prefetches data based on anticipated demand and analyzes patterns of data movement and task execution to inform future scheduling decisions. Our system significantly improves the processing times of HPC workloads by hiding data transfer time and employing high-quality, learning-based scheduling decisions. We evaluated our system with two different workloads: machine learning model training and image processing. We conducted performance comparisons using conventional data retrieval methods and the default Ray scheduler under various network conditions and storage configurations. Our findings consistently show that our system significantly outperforms traditional methods in every tested scenario.

Graduation Semester

2024-05

Type of Resource

Text

Handle URL

https://hdl.handle.net/2142/124562

Copyright and License Information

Copyright 2024 Yicheng Lu

Owning Collections