Page 126 - ITU Journal, Future and evolving technologies - Volume 1 (2020), Issue 1, Inaugural issue
P. 126
ITU Journal on Future and Evolving Technologies, Volume 1 (2020), Issue 1
processing), the O-RAN Distributed Unit (O-DU) is perform policy management, ML model
in charge of the High-PHY layer processing (e.g. management (described below in more detail) and
modulation, channel coding), Medium Access delivery of enriched information for near-RT RIC
Control (MAC) and Radio Link Control (RLC), the O- operation (e.g. RAN data analytics that could be
RAN Central Unit - Control Plane (O-CU-CP) hosts exploited by the near-RT RIC). Furthermore,
the upper layers of the control plane radio protocol complementing the A1 interface, the interactions
stack, i.e. Radio Resource Control (RRC) and control between the SMO and the underlying RAN nodes
plane of Packet Data Convergence Protocol (PDCP), also rely on the adoption of other standardized
and the O-RAN Central Unit - User Plane (O-CU-UP) interfaces named as O1 and O2 in Fig. 1. In
handles the upper layers of the user plane protocol particular, O1 refers to the set of service-based
stack, i.e. Service Data Adaptation Protocol (SDAP) management interfaces being standardized by
and user plane of PDCP layers. Then, sitting on top 3GPP for configuration, performance and fault
of these RAN nodes handling the distributed radio management of the RAN functionality [41]. In turn,
protocol stack, there is the near-real-time RAN the O2 interface supports the management of the
Intelligent Controller (near-RT RIC), which serves cloud infrastructure and resources allowing the
as the brain of the RAN by coping with the different execution of virtualized RAN functions.
Radio Resource Management (RRM) functions Building upon such a RAN reference architecture,
needed for overall RAN operation, such as radio Fig. 2 shows the main components and relations
connection, mobility, Quality of Service (QoS) and being delineated under O-RAN for the training and
interference management. With respect to the deployment of ML-assisted solutions within the
interfaces between these RAN nodes, E1, F1-c and SMO layer and/or within the RAN nodes themselves.
F1-u interfaces are specified by 3GPP while Open
fronthaul and E2 are being specified by the O-RAN ML Training Host
Alliance. Off-line data for
model training
ML Training
Service management and orchestration (SMO) (and evaluation)
Non-RT RIC
ML model deployment
A1 O1 ML Inference Host
Online data for
O2 model inference
Near-RT RIC
ML Inference
O-CU-CP Outputs
E2 (RRC, PDCP) E1 O-CU-UP Data Collection Actor
(SDAP, PDCP) and pre-
F1-c processing
F1-u
ML-Assisted
O-DU: RLC/MAC/PHY-high Solution
Open FrontHaul
O-RU: PHY-low / RF Data: Actions:
Management and operational -Internal actions within the actor
data (e.g. performance (i.e. subject of action is the actor itself)
O-Cloud measurements, number of -Configuration management over O1
connected devices, alarms) -Policy management over A1
Fig. 1 – O-RAN functional architecture from RAN nodes and UEs -Control actions over E2
Moving at the management plane, O-RAN defines
the Service Management and Orchestration (SMO) Fig. 2 – Components and relations for ML-assisted solutions
layer, which actually represents the Operations within O-RAN
Support Systems (OSS) of the MNO for the RAN As shown in Fig. 2, a variety of management and
domain. As part of the SMO layer, O-RAN basically operational data is collected from the different RAN
defines the role of a non-real-time RAN Intelligent nodes and User Equipment (UE) devices. Such data,
Controller (non-RT RIC) entity for the interaction properly preprocessed, is used to feed the two key
with the near-RT RIC via the A1 interface, which is components of the ML processing workflow,
also being standardized by the O-RAN Alliance. denoted as the ML training host and the ML
Through the A1 interface [40], the non-RT RIC can inference host. The ML training host represents the
106 © International Telecommunication Union, 2020