Predictive Closed-Loop Service Automation in O-RAN based Network Slicing

Compared to the rigid legacy wireless architectures, fifth generation (5G) technology promises to provide customized networks for vertical industries (like automotive, healthcare, agriculture, city management, manufacturing, etc.) with diverse service requirements. These diverse vertical services bring about a wide range of performance requirements in throughput, capacity, latency, mobility, reliability, position accuracy, etc. 3GPP Rel-15 standards [1] introduced the concept of Network slicing, for providing the foundation of a common connectivity platform for these different services and satisfying the diverse requirements of these services. Network slices are logical networks that are provisioned with a set of isolated virtual resources on top of a shared physical infrastructure  [2]. Network slicing can be realized by creating logical network functions and decoupling them from physical infrastructure. Currently, 3GPP Rel-17 standardization [1] activities on RAN slicing envision to empower 5G network slicing as a tool that operators can use to open up new source of revenue by involving the application providers in customization of 5G wireless design, deployment and operation for better support of the applications.

With resources being deployed as virtual machines rather than physical hardware, automation of service assessment and resource orchestration are important. Automation enables immediate recognition and correction of undesired events over network slices, without explicit human intervention. This evolution of zero-touch service assurance and closing the loop between recognition and correction significantly reduces the cost of creating flexible services adapted to the needs of customers [3]. However, providing such agility in the slice deployment, along with heterogeneity in topology, creates challenges in resource orchestration and dynamic service configurations [4]. Hence, advanced techniques are required to achieve optimal resource utilization while satisfying the requirements of multitude of services.

Given the development of machine learning (ML) techniques, intelligence is becoming a viable option for handling the heterogeneity of 5G wireless, including network slices [5]. The telecom industry has already demonstrated the use of ML in many trial deployments. Interestingly, 3GPP standardization has recently decided to look at ML with two different lenses: (a) one for the network and (b) one for the radio interface. The new 3GPP Rel-17 Study Item (SI) on the applications of ML to 5G RAN is just completed, and Rel-18 Work Item (WI) for standardization is approved to start from 2022 [6]. A separate 3GPP Rel-18 SI on application of AI-ML for improving the efficiency of wireless channels and radio interfaces is also approved to start from 2022 [7]. Hence, the de-facto wireless standardization is already poised to embrace ML for future 5G/B5G wireless. On the other hand, the Open Radio Access Network (O-RAN), an emerging new Radio Access Network (RAN) architecture, which builds on two core pillars that are intelligence and open interfaces, introduces specific ML-based functions like the Radio Intelligent Controller (RIC) [8].

In this article we provide an ML-based resource-provisioning scheme for Network slicing in the O-RAN framework to prevent Service Level Agreement (SLA) violations. A concrete practical example is discussed to explain details on the deployment architecture and the end-to-end workflow. In this intelligent scheme, AI techniques are used to perform predictions of future SLA violations and perform corrective actions in advance. Specifically, a recurrent neural network model is utilized to predict the amount of resources required over each slice, given the volume of traffic it carries. End to end O-RAN setup has been used for evaluation of the intelligent closed control loop, resource provisioning scheme, for Network slicing and control the physical resources and the cloud resource of the slices, respectively. This closed control loop operation continues until the defined SLAs are met. It is worth clarifying that the proposed O-RAN based intelligent and automated setup is the first proposal of it’s kind that can help in dynamically optimizing the over-served and under-served users in a slice, as the current implemented approaches in real-world networks are static solutions.

5G marks the transition to a more dynamic network architecture by introducing the concept of slicing. While the exact implementation of network slicing is operator specific, standards development organizations (SDOs) have provided template frameworks to facilitate this implementation ¹¹1ORAN Alliance, “O-RAN working group 1: Study on O-RAN slicing”, Tech. Rep., Oct. 2020.. A general framework of slicing implementation consists of two main phases i) slice instantiation and ii) resource orchestration over slices. During the first phase, a suitable set of network functions are configured to create a certain slice based on the defined slice specifications. Subsequently, the desired infrastructure resources for these network functions are allocated to meet the slice SLA requirements [9]. The isolation of resources on each slice is the key in preventing SLA violation between the slices. In general, there are two approaches for slice isolation: static and dynamic. In the static slicing, full isolation of the resource is guaranteed, albeit at the cost of possible over-provisioning of the resources. On the other hand, in a dynamic scenario, resources are multiplexed among the slices, by using advanced resource sharing and provisioning techniques.

As networks are becoming more dynamic and the resources are more heterogeneous, ML is needed to handle optimization of the networks and resources. It enables automation of resource modification and adaptation without human intervention. An automated network is based on a closed-loop control with adaptability enabled through feedback [10]. Many initiatives are focused on automated resource management for network slicing by using open platforms and open source tools, with the objective to provide automation across potentially multi-tenant slices and simplify resource orchestration [11]. Given that the O-RAN architecture is based on the principles of openness and intelligence, it can assist in providing predictive and closed-loop service automation by intelligently and dynamically orchestrating resources over the slices and maintain resource isolation. O-RAN architecture aims to reduce the proprietary implementation of hardware and software which in turn helps in increasing operational savings by establishing open standards. Additionally, openness of RAN components expedites the provisioning of new services for users. Moreover, intelligent RAN also reduces human intervention in the loop and increases the accuracy to handle the network complexity. The following are the main functional entities and interfaces (Fig.  1):

1. 1.

   Non-RT RIC: The Non-Real Time (Non-RT) RIC (Radio Intelligent Controller), which is part of the Service and Management Orchestration (SMO) system, provides a platform to deploy optimization applications, called rApps.

2. 2.

   Near-RT RIC: The Near-Real Time (Near-RT) RIC, which is deployed closer to RAN nodes, provides another platform to deploy optimization applications, called xApps.

3. 3.

   E2 Nodes: These nodes correspond to infield deployed RAN nodes supporting O-RAN standardized interfaces, like E2 and O1. E2 interface is between Near-RT RIC and E2 Nodes and is used by Near-RT RIC to configure E2 nodes based on policy formed at Near-RT RIC. O1 interface is between E2 nodes and the data collector. It is generally used by Service Management and Orchestration (SMO) Framework to collect RAN Performance Management (PM) statistics from E2 nodes.

4. 4.

   O-Cloud: O-Cloud corresponds to cloud computing platform where some of the functions of the RAN nodes and Near-RT RIC are deployed. There is an O2 interface between O-Cloud and SMO Framework. It is utilized to collect cloud statistics through data collector in SMO. It is also utilized to configure O-Cloud with the help of Service Orchestrator (SO) through policies formed at the Non-RT RIC.

5. 5.

   AI Server: AI Server hosts the platform for training machine learning models required for RAN optimization. For that purpose, we have utilized a dataset captured from E2 nodes of a real-world cellular network. The required data are collected through the data collector located in the service management and orchestration SMO layer. The intelligent prediction modelis trained in a proprietary AI server and deployed in the Non-RT RIC as rApp.

The two main components, to embed intelligence in O-RAN architecture and enhance the functionality of the traditional networks, are radio intelligent controllers, namely non-RT RIC and near-RT RIC. For the purpose of slicing, non-RT RIC collects key performance indicators (KPIs) related to slices, as well as the parameters that are used to configure them. These parameters and performance metrics are also sent to near-RT RIC to apply dynamic slice optimization for achieving SLA assurance. Based on the decision, network resources are controlled by near-RT RIC through the E2 interface and cloud resources by non-RT RIC through O2 interface ²²2https://www.o-ran.org. Please refer to [8], for detailed description of O-RAN goals, architecture and challenges.

With inspirations from software community we go a step ahead to integrate ORAN with real network components, along with our proprietary AI server, to form an inter-operable, end-to-end setup for testing dynamic resource provisioning in network slicing, using real world data. Our proposed RIC applications (i.e., Predictive Closed Loop (PCL) rAPP and PCL xAPP) checks the aforesaid network slicing use case in a closed loop automation environment. We believe this will further strengthen the O-RAN and network community to tackle stringent network slice resource management requirements, without the intervention of operators and prevent the SLA violation of the network slice(s). Our setup consists of wide range of Base Stations with variety of traffic patterns to make our solution more realistic and widely implementable. Fig. 2 represents the structure of deploying the proposed predictive closed-loop service automation.

Our real-world demo testbed includes open network automation platform (ONAP) and O-RAN software community projects hosted by Linux foundation that are publicly available to demonstrate the Predictive Closed Loop Automation. ONAP provides all the necessary functionalities related to service management and orchestration. The ONAP framework also provides an adapter to integrate proprietary AI server that is necessary to perform data training etc. This is expected to depict the Non-RT RIC functionality as listed in WG2 of O-RAN Alliance2. O-RAN SC provides all the necessary functionalities for Near-RT RIC, simulating the E2 Nodes (O-DU, O-CU etc.). The Cherry Release of the O-RAN SC is considered for the setup³³3 https://docs.o-ran-sc.org/en/latest. The network slices comprise of these E2 nodes. (Note that: O-DU (Open-DU) is the distributed unit that sits close to the radio unit and runs the Radio Link Control (RLC), Medium Access Control) MAC, and parts of the Physical (PHY) layer and O-CU (Open-CU) is the centralized unit that runs the Radio Resource Control) RRC and Packet Data Convergence Protocol (PDCP) layers. The information flow between components inside ONAP will flow through Data movement as a platform ${(DMaaP)}^{2}$ which is a premier platform for high performing and cost effective data movement services that transports and processes data from any source to any target with the format, quality, security, and concurrency).

1. 1.

   Considering the network slicing feature is not yet deployed by any operator in the world, we have mapped the existing 4G network data. Mapping of this real world data set to network slicing will be discussed in the next section. Field data collected from operator network is reported periodically via the O1 interface (defined in Section 2) to the Virtual Event Streaming ${(VES)\;collector}^{2}$ similar to that reported by an eNB in a commercial network data. Virtual Event Streaming (VES) Collector is a RESTful collector for processing JavaScript Object Notation (JSON) messages. (Please note that JSON is a text-based data interchange format to maintain the structure of the data). The related RAN statistics for each slice, including the number of active users per slice, volume of data and Physical Resource Blocks (PRB) utilization per slice are collecting in this manner. It is worth noting that there will be minimal processing for converting counters to KPIs and tag slice identifier (ID) at the data collector. Additionally the cloud resources related statistics are collected, such as virtual machines (VMs), computing units, memory units, from O-cloud (defined in Section 2) over O2 interface (defined in Section 2).

2. 2.

   The KPI information is sent to the proprietary AI Server which is installed on a different server where data cleaning, data training and data prediction happens.

3. 3.

   In order to learn the network or cloud usage pattern and predict the future values of the target parameters, an ML model is trained in the AI server. The trained model is utilized to predict the related KPIs which reflects the priority of the slice based on the network and cloud resource usage patterns.

4. 4.

   The outcome of the prediction is sent back to predictive closed loop (PCL) rAPP (rAPP defined in Section 2) that is running as an application in Non-RIC platform.

5. 5.

   PCL rAPP is an application host, part of Non-RT RIC. It also coordinates the formation of inference based on KPI prediction data and requirement of different slices based on slice level SLAs for network and cloud resources. PCL rAPP also prepares the RAN slice descriptor based on inference for configuring network resources. RAN slice descriptors contain slice IDs, public land mobile network (PLMN) ID, layer level individual descriptor and time stamp. The layer level network parameters that need to be scaled up/down in the individual layer slice descriptor and time stamps, can also be provided.

6. 6.

   PCL rAPP forms the prediction results for cloud resources and shares it with SDN-C/R which is responsible for configuration O-Cloud. (Please note that SDN-C/R is a Software Defined Network controller which is built on the common controller framework that assigns, manages and provisions network resources with the help of radio interfaces.)

7. 7.

   RAN slice descriptors will be passed on, over A1 interface by A1-controller, as a JSON policy to xAPP (xAPP defined in Section 2) using xAPP manager in Near-RT platform. The function of xAPPs applications is to provide programming capability for RAN component (O-CU/O-DU).

8. 8.

   xAPP will trigger the scaling up/down of network resources as defined in RAN slice descriptor by sending control message over E2 with the help of E2 Term and E2 Manager (inside Near-RR RIC platform). We define events in a way to change their parameters in E2 termination function (inside Near-RT RIC platform). Event definition is based on in the ASN format. It is a control message that specifies the event and state. Based on the proposed scheme, we require to define the new values related to PRB utilization and number of users per slice based on the prediction results prepared (by rAPP) with the help of predicted KPIs and their respective SLAs.

9. 9.

   On the other hand, O2 Client (inside SDN-C/R) helps in configuring (scale up/scale down) the cloud resources directly through O2 interface based on the inference output using O2 server in O-Cloud. If, due to prediction in a particular time instance, the PCL rAPP decides that the particular service is of less priority, it can bring down the cloud resources of a slice via O2 interface with the help of SDN-C/R. The deactivated cloud resources can also be activated based on the requirement of the service using same interfaces.

We now introduce our intelligent and automated resource provisioning scheme tailored for the O-RAN slicing architecture. Our proposed scheme learns temporal patterns of the traffic over four different traffic bearers with different service requirements by utilizing a deep learning based neural network. Such a learning is used to predict the number of users, volume of data and resource utilization for different cells in the network. To prevent a potential SLA violation, network resources are re-allocated accordingly. In fact, the goal is to utilize a commercial off-the-shelf (COTS) learning approach and incorporate intelligence in O-RAN architecture to proactively predict the resource requirements for efficient slicing.

Similar to legacy wireless networks, in 5G, standardized QoS identifier (5QI) are specified for bearers carrying different type of traffic corresponding to different types of services. The corresponding values of QoS to each identifier are provided in [12]. Therefore, bearers with different 5QI can be considered as RAN slices, as they are carrying traffic of different service types. Therefore, bearers with different 5QI can be observed as a way of representing a slice given the limited implementation of slicing in real-world networks as of today.

This article provides an intelligent closed-loop SLA assurance scheme for O-RAN slicing. The O-RAN architecture enhances traditional network function with embedded intelligence and open and standard interfaces. Building on the two core principles of openness and intelligence, O-RAN facilitates agile deployment and fast optimization of the slices that cater for, potentially, different verticals. We develop an intelligent and automated SLA assurance scheme through proactively re-orchestrating the resources over different slices. To this purpose, we utilize LSTM neural network to predict the number of users, volume of data and resource utilization on each slice. The effectiveness of the predictor model was demonstrated using a real-world traffic dataset that is taken from a cellular network in India. However, feedback loop to the ML model considered is under investigation and presupposed to be our future work. Based upon the predicted volume of traffic and resource utilization of each slice, available resources are dynamically adjusted to handle an under-served slice. The proposed scheme is carried out in a closed-loop and automated fashion without human intervention. A high-level architecture and the corresponding flow are also provided.

Utilizing the proposed scheme on top of O-RAN architecture brings in operational savings as a result of incorporating intelligence and automation. Operators can mitigate a lot of problems such as congestion, mobility management, etc. well ahead of time. Adoption of O-RAN gives an opportunity to the operators to integrate different intelligent solutions from different providers and create more competition. This subsequently brings down the capital and operational expenditures.