Wang, Haibo

Group Research Activities

To explore our new research idea in post network traffic measurements mining. The measurements are conducted online on network processors and then compressed and offloaded to offline server. Our purpose is to leverage advanced ML models to recover the measurements from the compressed measurements as much as possible, hoping to dramatically improve the performance.

Previous non ML network monitoring is to work on the high-speed high-rate network packet streams with limited online resources. This suffers from poor performance as information is compressed and lost. Only partial, approximate information can be kept.

Previous ML network monitoring is to directly feed the online packet streams to ML model and expect a good monitoring output. This is hardly done as the packet header in each packet can hardly provide enough information, especially statistical information.

What we do is to leverage the non-ML network monitoring method to fast process the packet stream and use ML model to learn on the compressed measurements. This “compress and learning” process can satisfy the limitation of online resources and maximize the potential of ML offline.

Projects

Step 1, using one basic non-ML monitoring method to output the measurements as input of the ML models. What we do on HPC is to feed the input to diverse ML models to find out a suitable one.

Step2. we will expand to different non-ML monitoring method as input.

Step 3, jointly optimize the design of non-ML monitoring and the ML models.

Computational Methods

What we need is GPU resources and we will explore current mature ML models to train our data

Group Members

Aayush Karki
Jiamu Song