Resumen de Data-Driven Self-Management of Cellular Radio Access Networks
Carolina Gijón Martín
In current years, cellular networks are experiencing profound changes to cope with the increasing demand for ever-diverse and ever-demanding services. As a result, the size and complexity of these networks has increased dramatically, evincing the need for zero-touch network and service management solutions. In the Radio Access Network (RAN), operators have already tackled the automation of management procedures in the past, giving rise to Self-Organizing Networks (SON). However, classical SON solutions are expected to be ineffective in next-generation networks offering services with extremely stringent and varying performance requirements. With the latest advances in information technology, it is now possible to leverage massive data collected in the Operations Support System (OSS) to develop advanced data-driven SON tools able to capture the peculiarities of each particular network. These new SON solutions must consider new features arising in 5G. One of these features is network slicing, allowing the coexistence of several separate logical networks operating simultaneously over the same physical infrastructure.
This thesis tackles the creation of data-driven self-management solutions for the RAN. Among existing SON use cases, the scope of this work focuses on two particular well-known self-planning and self-optimization use cases, namely RAN redimensioning and mobility load balancing. In both cases, solutions are proposed for legacy RANs, where all users share resources, and for new sliced RANs arising in 5G.
Regarding RAN redimensioning, this work explores the use of supervised learning over network data to derive performance models to detect potential capacity bottlenecks with radio planning tools. Models have been built for two purposes: estimating radio throughput metrics per cell/slice in different radio access technologies and forecasting cell traffic in the long term (i.e.,~months horizon).
Moreover, this thesis proposes two data-driven service-oriented mobility load balancing algorithms through handover parameter tuning. The main goal is to relieve local congestion issues by sharing traffic with neighbor cells. In both proposals, traffic steering has been formulated as a control problem. The first algorithm deals with traffic steering among cells in different carriers with quality of experience criteria, whereas the second algorithm tackles slice-aware traffic steering to guarantee service level agreement compliance in new 5G sliced RANs.
It should be pointed out that service-oriented self-management solutions proposed in this thesis require prior knowledge of the application demanded per user. However, obtaining such information nowadays is not straightforward for operators due to traffic encryption. The task of classifying encrypted traffic per service type is also addressed in this work. Such a problem has been tackled through unsupervised learning over connection traces, circumventing the need for a labeled trace dataset or the installation of expensive probes in the core network.
All the solutions proposed in this thesis rely on data currently available in the OSS, thus requiring no change in network infrastructure. To support the significance of results, performance assessment is always carried out in a realistic environment, i.e.,~with data from commercial cellular networks when possible or with a simulation tool calibrated with configuration and performance data from live networks otherwise.
