@INPROCEEDINGS{Sacc2309:RLVNA,
AUTHOR="Antonino Angi and Alessio Sacco and Enrico Alberti and Guido Marchetto and
Flavio Esposito",
TITLE="{RLVNA:} a Platform for Experimenting with Virtual Networks Adaptations
over Public Testbeds",
BOOKTITLE="2023 IEEE International Mediterranean Conference on Communications and
Networking (MeditCom) (IEEE MeditCom 2023)",
ADDRESS="Dubrovnik, Croatia",
PAGES=6,
DAYS=3,
MONTH=sep,
YEAR=2023,
ABSTRACT="Network emulators and simulation environments traditionally support
computer networking and distributed system research. The continued use of
multiple approaches highlights both the value and inadequacy of each
approach. To this end, several large-scale virtual networks testbeds, such
as GENI and CloudLab, have emerged, allowing testing of a networked system
in controlled yet realistic environments, focusing in particular on
facilitating the test of network management schema in Software-Defined
Network (SDN) scenarios. Nevertheless, setting up those experiments first
and integrating machine learning models later in these deployments is
challenging. In this paper, we propose designing and implementing a
web-based platform that integrates Reinforcement Learning (RL)-based models
with a virtual network experiment using resources acquired within a
real-world testbed, e.g., GENI. Users are able to reserve the network
resources (links, switches, and hosts) and configure them through our
intuitive interface with little effort. The RL algorithm is then launched
to learn how to steer traffic dynamically and according to diverse traffic
network conditions. Such a model can be easily customized by the user,
while our architecture enables fast reprogramming of the Open Virtual
Switches via the SDN controller instantiated. We experimented with
trace-based traffic to validate this user-friendly platform and evaluated
how centralized and decentralized RL algorithms can effectively lead to
self-driving networks. While in this paper, the system focuses on the
deployment of experiments for virtual network adaptation, the platform can
be easily extended to other network management mechanisms and machine
learning algorithms."
}