Soft real time resource management for vision based internet of things systems

Yang, Zhe

Permalink

https://hdl.handle.net/2142/117767 Copy

Description

Title

Soft real time resource management for vision based internet of things systems

Author(s)

Yang, Zhe

Issue Date

2022-11-21

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

Nahrstedt, Klara

Doctoral Committee Chair(s)

Nahrstedt, Klara

Committee Member(s)

• Godfrey, Brighten
• Xu, Tianyin
• Li, Baochun

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Internet of Things
• Computer Vision
• Computer Systems
• Machine Learning

Abstract

Vision based IoT systems have become ubiquitous today. These systems capture vision data – including images and videos – on cameras and offload data to an edge server or a cloud server for processing. Since cameras produce vision data continuously and the amount of produced data is very large, these vision data are usually consumed by automatic processing procedures including traditional processing techniques and deep learning based processing. Vision based IoT systems have low response time or soft real time processing as their service requirement. However, it is challenging to deliver soft real time services for these systems. First, when multiple cameras are deployed in nearby areas where they contend for network bandwidth, how to schedule network resources for cameras considering bandwidth, latency, and accuracy constraints? Second, when vision data are offloaded to edge or cloud and request to be processed by deep learning based computer vision models, how to schedule the processing of these tasks to deliver soft real time services? Third, for more complicated processing such as workflow processing, how to allocate resources in order to optimize processing response time? This thesis provides soft real time services for vision based IoT systems by tackling the above three challenges. Our thesis statement is: In vision based IoT systems, machine learning - system co-design is necessary to deliver soft real time services. We prove this statement by presenting a framework providing soft real time resource management for vision based IoT systems. This framework has three facets. First, we present a network bandwidth efficient streaming system which schedules network resources for vision based IoT applications when they offload vision data to edge or cloud to be processed by computer vision models. Machine learning based video processing techniques are employed to extract semantic information and reduce the amount of data to offload. With our ML - system co-design, we improve streaming utility by up to 23% compared to baselines when providing soft real time video analytics. Second, when vision data are offloaded to edge or cloud, a heuristics based soft real time GPU scheduler is presented to schedule the execution of computer vision workloads. With the ML - system co-design, our scheduler is able to greatly decrease deadline miss rates for computer vision applications run on the edge GPU. Third, as for more complicated processing involving workflows, we present a reinforcement learning based resource management framework to optimize workflow request response time. The ML - system co-design provides low response time services for complicated workflow processing on the cloud.

Graduation Semester

2022-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Zhe Yang

Owning Collections