Recent years witnessed a trend of “softwarization” of network components. Instead of static, expensive hardware, operators have started to adopt a more flexible approach based on Virtual Network Functions. This paradigm (aka Network Function Virtualization) advocates implementing network middleboxes such as firewalls or NATs as pieces of software to be deployed and executed on commercial off-the-shelf (COTS) hardware. This has boosted the development of several packet processing frameworks and software switches, which show nowadays multi 10-Gbps capabilities in COTS servers. In parallel, network systems are increasingly adopting machine learning (ML) techniques to solve complex networking tasks such as traffic classification or resource allocation.
As ML techniques require a large amount of data to be collected for both training and validation, when done in software, such measurements can highly affect the measured values, thus biasing the collected data. The intensity of this becomes stronger when measurements are taken close to the data path. Second, even after the training phase, complex model calculations may require dedicated hardware such as external GPUs or custom hardware designed for neural network processing such as TPUs or VPUs.
In this project, we present a novel approach based on non-invasive data collection relying on pure software. Our methodology consists of three components :
- low-impact network measurements with both direct and indirect observations
- inference/predictive modeling of a complete system with ML and/or classical approaches
- deployment of low-resource models for runtime query/action operations and automated recovery.