Towards AI-driven networks : hardware and software for data-plane ML
Abstract/Contents
- Abstract
- Network management is an increasingly difficult task for researchers and industry alike. Networks are growing rapidly in both scale and complexity. They now have to cater to a bigger application set and a larger user base than ever before while adhering to more and more stringent performance requirements. With so many challenges to running a network, operators must move beyond the era of hand-tuned algorithms and instead, adopt more automated approaches, i.e. AI-driven networks. In the search for more versatile tools in networks, many researchers have looked to machine learning (ML) as a vehicle for data-driven, adaptive mechanisms in networking systems. However, a number of pragmatic issues have plagued such development. Can we run ML in the packet path? Must operators build each new ML model by hand? How can we incorporate new data? In this dissertation, we show the construction of integral components required to build AI-driven networks. We first describe the design of Taurus, a platform to enable data-plane ML to run in the packet path of the network with a per-packet granularity, at line-rate. Furthermore, we demonstrate that the hardware for Taurus adds minimal overhead, less than 4% chip area and less than 3% power in our prototype. Next, we discuss Homunculus, a compiler stack for data-plane ML platforms (like Taurus) that allows for the automatic generation of resource and performance compliant ML models which outperform hand-tuned models by up to 16.9% in our tests. Finally, we show how these tools can be assembled to enable an adaptive ML loop in a network. Online labelling of raw data in the network can feed Homunculus, enabling the network to build new ML models from its own packet data. These models can then deploy learned policies in Taurus, setting the groundwork for upcoming AI-driven networks.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2023; ©2023 |
| Publication date | 2023; 2023 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Swamy, Tushar |
|---|---|
| Degree supervisor | Olukotun, Oyekunle Ayinde |
| Thesis advisor | Olukotun, Oyekunle Ayinde |
| Thesis advisor | Shahbaz, Muhammad |
| Thesis advisor | Winstein, Keith |
| Degree committee member | Shahbaz, Muhammad |
| Degree committee member | Winstein, Keith |
| Associated with | Stanford University, School of Engineering |
| Associated with | Stanford University, Department of Electrical Engineering |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Tushar Swamy. |
|---|---|
| Note | Submitted to the Department of Electrical Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2023. |
| Location | https://purl.stanford.edu/yk907xh7589 |
Access conditions
- Copyright
- © 2023 by Tushar Swamy
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...