Scientists:	Dr. Leonardo Linguaglossa, Dr. Fabien Geyer, Prof. Dr.-Ing. Georg Carle
Duration:	01.04.2019 – 01.04.2020
Funding:	German-French Academy for the Industry of the Future
Homepage:	https://www.industry-of-the-future.org/ai4perf/

Abstract

High-speed software networks are currently a hot topic in both industry and academia. Thanks to the tremendous advances in the techniques of software packet processing, the performance of software networks are now comparable to those of hardware middleboxes. This reduction in the performance gap comes with all the flexibility of software solutions as opposed to the purpose-specific nature of the hardware counterparts. Therefore, paradigms such as Network Function Virtualization (NFV) or Software-defined Networking (SDN) are becoming commonplace.

Today's ecosystem of frameworks for high-speed packet processing is extremely rich: evaluating the performance of such systems is important and challenging. One of the main difficulties is to obtain qualitative and quantitative that can objectively measure the performance. The project "AI4P: Artificial Intelligence for Performance." aims for developing a smart approach for performing tests and evaluating them automatically, thereby collecting meaningful data sets, and process them using advanced methods including machine learning algorithms. This exploratory project builds on expertise of IMT and TUM partners, linked together within the framework of the German-French Academy for the industry of the future.

The goal of the project concerns the analysis, modeling and evaluation of software network solutions, by proposing a new methodology for performance analysis and an experimental platform for the data collection, model validation and deployment of new network architectures. Our exploration focuses on different levels of the packet processing stack (from the bottom infrastructure, up to the network layer to the final provided service) as well as three orthogonal planes (consisting of the typical activities of monitoring, evaluation and optimization). We aim for targeted and reproducible measurements in combination with proper modeling of the complete system for improving performance in the relevant metrics.

As examples of the possible outcomes of the project we cite:

• What-if analysis for predicting the impact on performance in case of changes such as increased number of users, deployment of additional or more powerful devices or virtual machines, modification of routing, or update of virtualization technology;

• Root-cause analysis and bottleneck discovery in order to find where changes can be made in order to improve performance in the relevant metrics (e.g. Quality-of-Service / Quality-of-Experience);

• Detection of performance anomaly and correlation with misconfigured or misbehaving devices.

Partners

• Telecom ParisTech
• Airbus
• Cisco