The Leading Role of AI in the O-RAN Revolution
In the 5G era, capital expenditure (CAPEX) and operating expenditure (OPEX) are dramatically increasing. Driven by exponentially increasing traffic demands and with mobile networks pressured to become more energy efficient, changes need to be made to ensure networks adapt to these challenges whilst retaining high quality of service.
Enter the O-RAN Universe
In this context, the Open Radio Access Network, or O-RAN, paradigm is gaining steady traction. This industry-wide initiative advocates building and operating networks based on the principles of openness, intelligence, virtualisation, and automation. Interoperability is at its core, with open interfaces enabling different vendors to collaborate and innovate together.
At the heart of the O-RAN architecture is the RAN Intelligent Controller (RIC), which is a platform dedicated to the management of radio network and radio resources. The RIC is divided into a Non-Real-Time component (Non-RT RIC), which operates in the management plane and a near-Real-Time RIC (near-RT RIC) operating in the RAN domain. The Non-RT RIC operates with time granularity in the range of tens of seconds, minutes, or hours, while the near-RT RIC takes actions within tens or hundreds of milliseconds. Both RICs use external applications, named rApps and xApps, respectively. These [x/r]Apps include algorithms for use cases such as radio resource management (RRM), traffic steering, network slicing, etc. xApps and rApps can include AI/ML logic and are key enablers for autonomous networks and AI-driven RAN resource management.
Managing and Optimising Resources with xApps
xApps are deployable in the near-RT RIC and hence operate in near-real-time. To achieve the highest operational efficiency in pursuit of automating network management, the xApps should be cloud-native microservices that ingest relevant performance measurement data from different network elements, e.g. eNBs or Centralised Units (CUs), via the standardised E2 interface. Based on this data, the xApps’ decision logic will return use-case specific commands, e.g. allocation of compute capacity, on/off switching of radio resources, etc. back to the relevant nodes.
Net AI’s Resource Autoscaling xApps Powered by Artificial Intelligence
One of the problems we tackled recently at Net AI is the automatic allocation of the right amount of resources at the CU (Centralised Unit) level when they are needed, thereby meeting customer demands in a cost-effective way, while reducing operational costs.
To do this, our xUPscaler xApp embeds our proprietary AI models and uses historic and real-time network traffic data to predict upcoming traffic volumes at the level of base stations (or at higher geographic granularity). We use these forecasts along with our AI-driven autoscaling logic to compute relative “gNB-CU-UP Capacity” values, which can then be used to load-balance traffic between gNB-CU-UPs (gNB Centralised Unit User Plane). This allows scaling of network resources ahead of time and more efficient resource management. You can watch a demo of xUPscaler in the following video:
Under the Hood
In the scenario designed for the above demo, one or more O-RAN E2 nodes (here CUs) send Key Performance Metrics (KPMs) to the near-RT RIC via the E2 interface. The KPM we consider in our use case is aggregate traffic volumes for each cell in a target geographical area.
Our xApp then subscribes to this KPM.
The xApp embeds Net AI’s SoothsAIer forecasting engine, which exploits both spatial and temporal correlations that are characteristic to mobile traffic, to accurately predict future traffic volumes at each cell. These forecasts are also provided to the near-RT RIC Shared Data Layer (SDL), so that they are available across the entire RIC and can be used by other native functions or third-party xApps. On the other hand, our xApp also aggregates the antenna-level predictions at the level of the gNB-CU-UP, and feeds those aggregates to a second embedded AI engine, which performs an anticipatory auto-scaling of computational resources by generating a relative capacity allocation for each gNB-CU-UP.
The gNB-CU-UP capacity allocation messages are finally sent by our xApp to the E2 node through the E2 interface to scale network resources ahead of time and handle cell load more efficiently.
Where to Next?
In a recent post we spoke about the growing importance of improving the energy efficiency of mobile networks and the instrumental role AI can play. We recently demonstrated how accurate traffic forecasts obtained with our AI technology can be used to control the number of carriers that are live in the RAN at any point in time, so that the energy consumption can be significantly reduced, while minimising the number of under-provision events, which impact capacity and therefore user quality of experience. More details about that in a future post.
Further integration and testing is currently underway with RIC platforms developed by different vendors. We remain committed to helping communication service providers reduce the carbon footprint of their 5G infrastructure through the application on AI driven intelligence and automation.
