Metis: Understanding and Enhancing In-Network Regular Expressions

Authors: Zhengxin Zhang, Yucheng Huang, Guanglin Duan, Qing Li, Dan Zhao, Yong Jiang, Lianbo Ma, Xi Xiao, Hengyang Xu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Metis is more accurate than original REs and other baselines, achieving superior throughput when deployed on network devices. We collect network traffic data on a large data center for three weeks and evaluate Metis on them. 4 Experiment 4.1 Experiment Setup
Researcher Affiliation Collaboration Tsinghua University, Peng Cheng Laboratory, Northeastern University Tsinghua Shenzhen International Graduate School, Tencent
Pseudocode No The paper describes various algorithms and processes (e.g., Thompson's construction, DFA construction, SSKD steps) but does not include structured pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Our codes are available at Github2. (Footnote: https://github.com/You Are Special To Me/Metis.)
Open Datasets No The paper states that "we collect packet-level traces at gateways of a large data center" for evaluation, indicating a custom dataset that is not explicitly made publicly available with specific access information (link, DOI, formal citation).
Dataset Splits Yes We split each category of the dataset into the training set, test set, and validation set with a ratio of 7 : 2 : 1.
Hardware Specification Yes We conduct our experiments on a server with two 16-core CPUs (Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz), 64GB DRAM memory, and six Ge Force RTX 2080 SUPER GPUs.
Software Dependencies No The paper mentions software like "P4 language", "Barefoot P4 Studio Software Development Environment(SDE)", and "Intel DPDK library", but does not provide specific version numbers for these components.
Experiment Setup Yes For DFA2BRNN, we set the learning rate to 10 4, batch size to 500, and hidden state to 200 for each model. We use the cross-entropy loss as the objective function. We train each model for 200 epochs and use early stopping to avoid overfitting. For BRNN2PSRF, we set t to 7, α to 0.3, # split features to 7 and minimum samples to 15.