24
AMF UPF
PCRF HSS
SMF
RRU BBU
UPF
PCRF HSS
SMF
RRU BBU AMF
AMF UPF
PCRF HSS
eNB SMF
RAN functions Core network Cloud Edge
Antenna
Slice A
Slice B
Slice C
Figure 2.12:VNF placement considering three different slices, and antenna, edge and cloud locations.
a mid-term in the relation between resources and costs. Slice A concentrates its VNFs in the cloud, which guarantee a lower cost in relation to slice B and C, but presents a greater latency to applications. Slice B moved some of core network functions to edge location to guarantee a decrease in the latency and slice B moved all core network to edge and the eNB to antenna location, which increases the costs but guarantees the best latency and applications response possible among these mobile network slices.
Other aspects besides latency and cost should be considered to make an efficient VNF allocation, such as an analysis of the physical resources available in each location and network link occupancy to better attend conditions defined on SLA [51].
25 Network slicing provides an opportunity to deliver mobile network services to attend different clients in a variety of scenarios and requirements, but it requires that mobile network be able to adapt their parameters to different conditions. Therefore, AI and more specifically ML methods are very important to explore network slicing into mobile network context.
The current mobile network infrastructure limits the potential ML benefits because it is not prepared yet to accommodate ML-oriented tasks [6]. To enable ML-aware networks is essential to define standards that allows different manufacturers and operators to develop tools and solutions to be integrated into a mobile network system. The FG-ML5G from ITU-T has specified a unified architecture for 5G and B5G that provides a common nomenclature for ML- related mechanisms guaranteeing interoperability among networking systems [9].
Fig. 2.13 shows the high-level framework proposed by FG-ML5G to integrate ML in future networks. ML pipeline is a set of logical nodes with specific functionalities which can be combined to build a ML application in a telecommunication network. This ML pipeline can be integrated to a simulated or a real ML underlay network that provides the network information to the pipeline and is the target of the ML operations.
SRC C PP M P D SINK
ML Underlay Networks ML Pipeline
3 Data handling reference point Data handling reference point 4
SRC C PP M P D SINK
Simulated ML Underlay Networks ML Pipeline
1 2
ML Sandbox
Data handling reference point Data handling reference point
ML Intent MLFO
Reference points among ML pipelines nodes ML intent input to MLFO
Service ingress point (consumer) Service egress point (producer) MLFO: ML Function Orchestrator
C: Collector PP: Pre-processor M: Model P: Policy D: Distributor SRC: Source of data SINK: Target of ML output
Figure 2.13: High-level architectural framework for ML in futures network based on ITU-T Rec.
Y.3172.
26 Source (SRC) component is the source of data to be used by ML pipeline. In a mobile network, it could be elements such as a UE, a RAN or a core network component. Collec- tor (C) component aggregates information from all SRC nodes. Preprocessor (PP) component makes the data pre-processing to prepare the information to ML model entries. Model (M) component represents the ML model, usually a neural network. Policy (P) component enable the application of policies to the output of the model node. Distributor (D) component identifies sink (SINK) components and distribute the ML model output. SINK component is the target of ML output, which usually applies actions suggested by the model in the mobile network to perform improvements in the network operations. Service ingress and egress point represents the entry and output of each component.
ML sandbox is an isolated domain with a simulation of a mobile network system, which enables to train, test and evaluate the ML operations before applying them in a real mobile net- work. The ML function orchestrator (MLFO) is a logical node that manages and orchestrate the components of ML pipeline based on ML intents and network conditions [9]. The MLFO func- tion receives the ML intents information through service ingress point and orchestrate the ML components over a NFV structure through a communication using the service egress point. This work implements MLFO funtions using a combination of Kubernetes [34] and Kubeflow [54]
softwares. Therefore, ML framework defined by FG-ML5G enables the integration of ML ap- plications to mobile network infrastructure, giving flexibility to these applications to adapt over distinct physical resources similarly to NFV system. In this work, it is used a combination of softwares Kubeflow [54] and Kubernetes [34] to implement this AI/ML architecture using a virtualized and orchestrated functions over a cluster of machines.
Next chapter starts a discussion about the CAI testbed components implementation, fo- cusing on implementation details as the virtualization, orchestration and connected AI structures used to generate the mobile network to perform tests.
Chapter 3
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 dynam- ically 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. The ML workloads are orches- trated along the cluster to provide the AI agent processes. All the code to set up the testbed and execute the use cases, which will be explained in the following sections, are available on Github.1 The next sections detail each aspect of the testbed.