FCoder: A real-time large-scale bottleneck detection mechanism with neural network and transfer learning

Zhu, Zhoushi

Permalink

https://hdl.handle.net/2142/110867 Copy

Description

Title

FCoder: A real-time large-scale bottleneck detection mechanism with neural network and transfer learning

Author(s)

Zhu, Zhoushi

Issue Date

2021-04-27

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

Hu, Yih-Chun

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Shared bottleneck detection
• Machine learning
• Transfer learning
• Siamese neural network
• TCP Multipath

Abstract

Detecting shared bottlenecks among network flows is crucial in TCP Multipath to ensure TCP fairness and other applications related to cross-flow congestion control. The problem of whether two flows share a bottleneck has been well investigated in previous work, but a large-scale bottleneck detection scheme among a large number of flows remains not fully explored. In this thesis, we explore how to scale the pairwise shared bottleneck detection mechanism to work with a large number of flows using machine learning techniques. We can use machine learning techniques to preselect possible candidate pairs before pairwise comparison. It improves the time efficiency, reduces waste of computational resources, and enables scaling to detect among a large number of flows, compared with pairwise examining all existing pairs of flows. To validate the idea, we present fCoder, a solution for large-scale shared bottleneck detection for Internet-like networks that detect co-bottlenecks using a correlation mechanism based on time-domain samples of the round-trip-time (RTT) of flows. We accelerate the process using machine learning techniques to enable detection among a large number of flows in real-time. We evaluate the detection performance and time efficiency of fCoder to demonstrate that Machine Learning is a promising technique that helps scaling shared bottleneck detection approaches to detect shared bottlenecks among a large number of flows. In particular, we have shown in our experiment that we can process 100 times more pairs of flows by training a Dense Neural Network (DNN) using simulated network traces data. We have also shown that we can use the transfer learning technique to tune the DNN using a small amount of real-world data to accurately detect real-world datasets.

Graduation Semester

2021-05

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/110867

Copyright and License Information

Copyright 2021 Zhoushi Zhu

Owning Collections