Graduate Thesis Or Dissertation
 

Predicting and Improving Throughput, Responsiveness and Battery Life of Computer Systems by Machine Learning

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/n870zz66n

Descriptions

Attribute NameValues
Creator
Abstract
  • The Machine Learning (ML) algorithms are increasingly explored in varies of fields including designing and optimizing computer systems. Recent research, such as optimizing memory/cache prefetching by ML training or predicting traffic pattern in throughput processors, also exhibits a promising future of introducing ML into computer system design and optimization. Throughput optimization in throughput-oriented processors is imperative as the computing workload of parallel and cloud computing have been growing rapidly in recent years. At the same time, throughput optimization can be time consuming when applying conventional design and optimizing process as the design space is prohibitively huge. In the first part of this dissertation, we firstly define a huge and complicated design space in silicon interposer-based throughput processors, then utilize Monte Carlo Tree Search model (MCTS) to exploit and explore the design space. The evaluation results show that the system performance is improved by over 20\% with only 0.05\% design space is assessed. Performance and power (PnP) are also the most important metrics that are utilized by original equipment manufacturers (OEMs) to conduct the design of a device. Current PnP measurements rely on manual hardware swapping and testing for systems which is time consuming and not financially-efficient. A fast and accurate PnP value prediction solution can guide OEMs to understanding the basic behaviour of different hardware components, and, more importantly, shortens the time to market of a device. In the second work, we explore the common ground between natural language processing (NLP) problems and system PnP prediction problems, and develop an NLP-like solution to resolve the problem. The solution is available to extract the inter- and intra-relationship among the existing system components and to predict the behavior of system components that have not appeared before. The results of our evaluation demonstrate that the solution achieves as high as 94\% labeling accuracy in a real-measured dataset.
License
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Embargo reason
  • Intellectual Property (patent, etc.)
Embargo date range
  • 2021-05-29 to 2021-12-30

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details LiYunfan2021.pdf 2021-05-29 Public Download