Reproduce “Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane” on OpenResearch

About this paper

Yinchao Zhang, Su Yao, Yong Feng, Kang Chen, Tong Li, Zhuotao Liu, Yi Zhao, Lexuan Zhang and 4 more

Pegasus is a framework that enables scalable deep learning inference directly on commodity programmable switch dataplanes, leveraging novel techniques to overcome architectural limitations. It achieves enhanced accuracy, greater scalability for model size and input dimensions, and wider generality across various deep learning models, facilitating real-time intelligent network functions.

Key insights

• The primary limitation for deploying DL on dataplanes stems from an "ineffective use of the MAT abstraction" rather than inherent resource scarcity.
• Fuzzy matching provides a scalable alternative to direct input-output mapping by leveraging learned cluster representations, effectively compressing data for efficient dataplane processing.
• Strategic primitive fusion can drastically improve resource efficiency and allow for the deployment of much larger and more complex DL models on constrained hardware.
• Fixed-point quantization, carefully applied with backpropagation for fine-tuning, can preserve high accuracy while adhering to dataplane numerical constraints.

Summary by alphaXiv.