Serverless Enabled Framework for Machine Learning Applications: Architecture, Solutions, and Evaluations

Authors

Boyuan Guan, Florida International UniversityFollow

Document Type

Dissertation

Degree

Doctor of Philosophy (PhD)

Major/Program

Computer Science

First Advisor's Name

Liting Hu

First Advisor's Committee Title

Committee chair

Second Advisor's Name

S.S. Iyengar

Second Advisor's Committee Title

Committee member

Third Advisor's Name

Mark A. Finlayson

Third Advisor's Committee Title

Committee member

Fourth Advisor's Name

Wensong Wu

Fourth Advisor's Committee Title

Committee member

Keywords

computer science, physical sciences and mathematics

Date of Defense

11-9-2022

Abstract

Over the past decade, the popularity of machine learning applications such as recommendation systems, image recognition, real-time alerts, event detection, natural language processing, and online streaming analytics has increased dramatically. However, there is a long-tail problem has been identified for the ML application. The majority of ML applications are concentrated in high-tech and high-profit areas while ML applications are still difficult to reach for local and low-profit businesses. The two factors that cause this slow growth are the domain problem barrier and the rigid infrastructure environment. Since ML application development is different than the traditional software application which consists of data steam feeding, data prepossessing, model training, evaluation, and service update. Unlike traditional software application architecture which the data layer is separate from the business layers, the data is involved in all the steps in the ML application development process. In other words, domain knowledge is required for the overall life cycle of ML application development. This requirement causes big overheads in the workflow and thus, produces the barrier to the widespread of ML applications in local low-profit businesses. On the other hand, the various types of ML applications and the heterogeneous infrastructure environment make the ML application beyond the reach of these local low-profit businesses. Compared to high-tech and high-profit companies, local low-profit businesses can not afford an IT team to implement their ML applications. Pure

cloud solution normally provides template-like services which are hard to adapt to their own business workflow. Even if they are able to implement some low-complexity ML applications within an on-premise private cloud environment, the application performance is hard to meet a production level stability.

In this dissertation, we present two frameworks: dpSmart and StraightLine aimed at solving the domain-specific problem and heterogeneous infrastructure problem for the ML application. dpSmart is a conceptual framework that provides an additional abstraction on top of the existing ML application development process. The purpose of the additional abstraction is to limit the domain-specific knowledge involvement to one single step so that the rest of the steps can be standardized. Straightline is a from-development-todeployment multiple resources-ware machine learning application pipelines. It first separates the ML application development and deployment phase so that various ML applications can be served without affecting the deployment environment. StaightLine adapts the docker container to provide flexibility for dynamic implementation among heterogeneous infrastructure environments. StraightLine also presents a placement algorithm to maximize the ML application performance.

Identifier

FIDC010976

Recommended Citation

Guan, Boyuan, "Serverless Enabled Framework for Machine Learning Applications: Architecture, Solutions, and Evaluations" (2022). FIU Electronic Theses and Dissertations. 5142.
https://digitalcommons.fiu.edu/etd/5142