RNS INSTITUTE OF TECHNOLOGY

Channasandra , Bengaluru-560061
Dept. of ELECTRONICS AND COMMUNICATION ENGINEERING
2022-2023

Technical Seminar on:

Testbed for 5G Connected Artificial Intelligence
On Virtualized Networks

Under the Guidance of: By:

Sreenivasa Babu M O, Syed Arham Iftekhar
Assist. Professor, USN: 1RN19EC151,
Dept. of ECE, RNSIT. Dept. of ECE, RNSIT.
 CONTENTS
1. ABSTARCT
2. INTRODUCTION
3. CONTRIBUTIONS
4. RELATED WORK
5. CAI
6. VIRTUALIZED FRONTHAUL AND BACKHAUL
7. RAN PROGRAMMABILITY
8. USE CASES
9. CONCLUSION
 ABSTRACT
• The fifth-generation (5G) cellular networks incorporate a large variety of
technologies in order to address very distinct use cases. Assessing these
technologies and investigating future alternatives is complicated when one relies
only on simulators.
• 5G testbeds are an important alternative to simulators and many have been
recently described, emphasizing aspects such as cloud functionalities,
management and orchestration. This work presents a 5G mobile network testbed
with a virtualized and orchestrated structure using containers, which focuses on
integration to artificial intelligence (AI) applications.
• The presented testbed uses open-source technologies to deploy and orchestrate
the virtual network functions (VNFs) to flexibly create various mobile network
scenarios, with distinct fronthaul and backhaul topologies. Distinctive features of
the testbed are its relatively low cost and the support to using AI for optimizing
the network performance.
• The paper explains how to deploy the testbed structure and reproduce the
presented results with the provided code. AI-based radio access network (RAN)
slicing and VNF placement are used as examples of the testbed capabilities.
 INTRODUCTION
• The fifth-generation (5G) cellular network addresses a variety of usage scenarios, including
enhanced Mobile Broadband (eMBB), Ultra Reliable Low Latency Communications (URLLC)
and massive Machine Type Communications (mMTC).
• In order to meet very distinct requirements, 5G networks rely on a flexible New Radio (NR)
interface and several virtualization technologies, such as Network Function Virtualization (NFV),
Software-defined networking (SDN) and Software Defined RAN (SD-RAN).
• Using virtualization, a better use of computational and network resources is achieved, with the
resources being allocated according to the applications’ requirements, enabling efficient strategies
such as network slicing . All the flexibility aimed by 5G to attend the large number of different
scenarios leads to an increase in the complexity in the network management.
• Artificial Intelligence (AI) tools The associate editor coordinating the review of this manuscript
and approving it for publication was Qichun Zhang . compose an important part of the strategy to
automatize the deployment and configurations of parameters in a 5G network, aiming at
achieving, for instance, self-organizing networks (SON) and zero touch networks .
• AI applied for mobile networks is a key enabler to manage not only 5G but also beyond 5G (B5G)
networks. The existing 5G mobile networks (3GPP Releases 15 and 16) have no sufficient
flexibility and intelligence to fulfill all requirements for the three main use cases (eMBB, URLCC
and MTC) yet . Over the years, the networks were not appropriately designed to accommodate AI-
oriented tasks such as data collection, processing, and output distribution.
• Several standardization bodies are working to incorporate AI / ML in 5G and B5G
mobile networks, such as the 3rd Generation Partnership Project (3GPP),
European Telecommunications Standards Institute (ETSI) and International
Telecommunication Union Telecommunication Standardization Sector (ITU-T).
• This paper describes the Connected AI (CAI) testbed structure to build flexible
and realistic scenarios with different network topologies for 5G and quickly
deploy them. A container-orchestrator makes the deployments of the applications
to create a mobile network interconnected to virtualized OpenFlow switches, an
SDN controller, and a RAN controller.
• The testbed is built with open-source software and Virtual Network Functions
(VNFs) to facilitate the deployment and research of different network scenarios
without the need to use specialized hardware.
• This paper presents two use cases, one for RAN slicing and another for the
placement of VNFs according to application requirements. The source code is
made available to allow replicating these two experiments. The experiments were
designed to be easy to follow, and decrease the learning curve to get started with
AI techniques applied to 5G using the CAI testbed.
 CONTRIBUTIONS
The contributions of this paper are the following:

• It presents a low cost testbed that emphasizes AI integration using Kubeflow

to the ML workloads associated to the management of 5G networks. The
proposed testbed avoids the complexity of Management and Orchestration
(MANO) agents to provide cloud services and focuses in AI applied to RAN
and transport network, including fronthaul and backhaul.
• The paper describes two use cases that explain how AI can be used in the
testbed, leveraging collected information and applying AI decisions to
optimize the network.
• The experiments are reproducible, given that the code to deploy the testbed
and the two use cases applications are made publicly available.
RELATED WORK: TESTBEDS AND THEIR FEATURES
• Several 5G testbed implementations have been developed to demonstrate functionalities
and operation of a 5G system. Given the variety of technologies in 5G, to make the
following discussion more concrete, special emphasis is placed on slicing and VNF
placement, which are the two use cases described in Section IV.
• An important feature of 5G networks is the possibility of an infrastructure (InP) to share
its physical resources among different Mobile Virtual Network Operators (MVNOs),
providing isolation, security, and accomplishing the Service Level Agreements (SLAs)
requested by each operator.
• NFV is essential to make better use of physical resources and making the VNFs
allocations along the infrastructure using orchestrators. The network slicing facilitates the
MVNO and InP provisions of different services in the network with dedicated resources.
• With the wide variety of applications and scenarios, it becomes more important to choose
the better network parameters and VNF placement to improve the performance of each
network slicing without affecting others.
CAI: THE IMPLEMENTED CONNECTED AI TESTBED

The CAI testbed enables the building of flexible and realistic AI-based
scenarios with different network topologies for 5G and quickly deploy
and assess them. The SDN and RAN controllers work as information
sources about the network. They also work as agents to dynamically
change the mobile and the computer network. An AI agent performs
different actions in the testbed according to the application, using the
information provided by SDN and RAN controllers to train and execute
in test stage its neural networks.
Fig. 1 shows the testbed modules interconnection considering a C-RAN scenario, which
is one of the scenarios available in CAI. The figure depicts a virtualized fronthaul
deploying one specific topology, which can be redefined with flexibility via scripts. Both
SDN and OAI FlexRAN controllers work as an information source to the AI agent.
Fig. 2 depicts the NFV architecture implemented in the testbed. It is similar to an NFV
architecture implemented with MANO, but simpler and with less functionalities. The
NFV Infrastructure (NFVI) is composed of the hardware resources which are
represented in the testbed by the computers of the Kubernetes cluster. The virtual
computer, storage and network are allocated using Kubernetes and Docker through the
virtualization layer, which allocates the virtualized components in the hardware
resources. The VNFs are represented by the containerized elements that compose the
VIRTUALIZED FRONTHAUL/BACKHAUL AND
SDN CONTROLLER

In a C-RAN deployment using RRU and baseband unit (BBU), the fronthaul is the link
connecting these modules. Similarly, the backhaul is the link connecting the BBU to the
core network. OAI implements both fronthaul and backhaul over an Ethernet
connection. Each scenario of mobile network deployment presents different topologies
to the fronthaul and backhaul represented by physical components built to enable the
transport network, e.g., routers and switches. The testbed implements both the fronthaul
and backhaul using a virtual network deployment based on Mininet. The use of Mininet
gives more flexibility to deploy the fronthaul and backhaul since any network topology
can be defined using Mininet scripts, which set virtualized routers and switches to
compose the network.
• The virtualized routers/switches created with Mininet cannot communicate
directly to networks outside the virtualization since it can see only the links inside
Mininet virtualization as defined in the topology script.
• To create a link between the host machines and Mininet hosts, we created virtual
Ethernet devices (VETHs) linking them, enabling the traffic forwarding between
two machines through the Mininet topology. Therefore, the traffic of the
fronthaul/backhaul can be forward through Mininet virtualized network.
• Figure 4 illustrates how the RRU and BBU machines forward their fronthaul
traffic through the virtualized topology instantiated using Mininet. This figure
does not show the SDN controller or other Mininet hosts (routers and switches)
because it focuses on how machine interfaces are connected to the virtual network
to facilitate understanding. Each machine has an Ethernet interface (ETH)
connected to the same network. The Mininet hosts can communicate with each
other through virtual interfaces (VIs) and virtual switches defined in Mininet
script topology, but they are not able to communicate with the ETHs from
machines, so the fronthaul traffic is not able to be forwarded through Mininet
topology using the default deployment.
 RAN PROGRAMMABILITY
• RAN programmability, also called SD-RAN, works as an abstraction of the RAN
resources and by providing an API which enables the Service Orchestrator entity to
dynamically manage the RAN resources to provide information about the mobile
network.
• The FlexRAN protocol has defined and implemented an SD-RAN architecture integrated
with OAI platform which incorporates an API to separate control and data planes for the
mobile RAN.
• FlexRAN APIs enable the development of applications related to the control and
management of the RAN resources , e.g., schedulers, interference, and mobility manager.
• Moreover, applications related to improvements in the use of RAN resources to make
more sophisticated decisions , such as RAN slicing and adaptative video streaming based
on channel quality. FlexRAN does not control the flows in the wired domain, so it does
not support the management of routes, packets filtering, and other functions related to the
computer network domain (routers and switches). These functions can be executed using
the Ryu controller.
 CAI TESTBED USE CASES

• The proposed testbed structure can be applied in different scenarios

and be integrated into different AI agents to improve the mobile
network functions. The next subsections provide two use cases of the
AI agent.
• The first is a RAN slicing application to monitor the percentage of
resource blocks (RBs) allocated for each requested slice.
• The second application is an AI agent to define the placement of VNFs
along with the testbed’s cluster machines.
 RAN SLICING
• An application for RAN slicing is proposed, utilizing an in-network deep learning
to analyse packets in the Packet Gateway (PGW) at the core network and
identifying mobile applications. Afterwards, the eNB can apply specific slice
configurations for each UE according to requirements associated to the detected
application.
• The work presented designs and prototypes a network slice solution in a C-RAN
architecture that aims to share spectrum among slices efficiently considering their
requirements.
• The authors use FlexRAN information to identify application requirements and
verify whether a service request can be satisfied according to the real-time
network state, and communicates the slicing decision. It uses an algorithm for
dynamic slicing that estimates the resource allocation satisfaction and shares
unused RBs among applications which requires more resources.
• We implemented a RAN slicing scenario using FlexRAN and considering eNBs in
a C-RAN architecture, where Mininet virtualizes the backhaul to implement a
network topology with switches and routers.
Fig. 6 shows the scenario with 3 UEs (smartphones) connected to the RRU.
A fronthaul composed by a single path connects the RRU to the BBU pool.
The BBU pool is connected to the Free5GC core network through a
virtualized backhaul using Mininet, containing two routers and one switch
connected to the Ryu controller. The total latency added in the backhaul was
100 ms. In this scenario, the main source of information is the Ryu SDN and
the OAI FlexRAN controllers. A customized script makes the Ryu controller
provide throughput, link delays and jitter in backhaul link, besides the
default information provided by the default Ryu API, e.g., number of packets
transmitted and received in each switch port. The FlexRAN provides
information about the eNBs, RAN slices and UEs connected, such as the
percentage of RBs allocated for each slice, resources isolation and the
identification of all UEs connected in a specific eNB.
 VNF PLACEMENT

• The next paragraphs discuss how the CAI testbed was used in a VNFs placement
application. CAI leverages the implementation of mobile network modules as
containers, which enable the deployment of different VNFs architectures over the
physical structure.
• For instance, CAI supports the test of different processes of network slicing. The
requirements of a requested slice, defined in the SLA, can be processed by a
function mapper, which uses the testbed to implement the VNFs in the locations
of the network to better fulfill the requirements.
• Similar to the RAN slicing use case, we adopted DNN based supervised learning
for VNF allocation, but CAI can support other ML and AI paradigms such as RL,
as well as take in account aspects such as energy consumption and network
congestion.
 CONCLUSION

• This paper presented a flexible 5G testbed with RAN and SDN controllers,
connected via virtualized backhaul and fronthaul.
• The main focus of the proposed CAI testbed is enabling ML-based
applications. When compared to other testbeds, CAI has the advantages of
being reproducible among distinct sites given the provided containers, and
lower cost than previously published alternatives.
• The authors are now investigating the Elasticsearch stack to provide the
storage of the controller’s information in a time series format, improving real-
time training of AI agents.
• Elasticsearch also facilitates gathering information from the computers used
in the testbed, such as load, CPU and RAM memory usage.
THANK YOU