Until a few years ago, mobile networks traffic had a predictable demand pattern. Hence, planned node placements or capacity assignments with network processing software capabilities tied to the particular physical node worked reasonably well. Designers could run networks near-optimally within the constraints using well-known Erlang models with given call blocking probabilities. The network state was often kept in near static simple repositories such as trees, tables etc.

However, in the last few years data usage has increased significantly. Data is much harder to model than voice, with new applications such as Pokemon GO being virally popular for just a few weeks before tapering off. Network designers tried to address new demand patterns via hardware-software separation using virtualization methodologies such as NFV with similar repository architectures as before. This method somewhat worked since the traffic was still human-centric. Recently, with the increasing power of data and data-analytics, “things” have also started to get connected. These “things” range from small meters and cookers to huge wind-turbines and factories, spitting out data at rates anywhere from a few bits to multiple gigabits every second. Orthogonally, there has been a growth spurt in new applications and business models, such as ride-sharing, Industry 4.0 etc. that fundamentally assume the presence of connectivity at every time and place with reliability requirement of 5 or 6 9s. With network having to address not only human-human communication but also that of human-to-machine, machine-to-machine, machine-to-cloud and more, a new network was needed. The new 5G standard is designed with precisely these requirements in mind - ability to reliably handle the extreme variability in requirements for all traffic types.

From the beginning, 5G was designed with flexibility in both radio interfaces and access and core networks. However, by itself, it was not sufficient since the same network was required to simultaneously serve increasingly diverse use cases. This led to a new cornerstone of a successful 5G network design: the concept of network slicing.  ETSI has defined a network slice as a description of a service aware logical network that is composed of different physical or virtual network elements, resources and functions, often for the purpose of efficient utilization of the network while meeting the required service specificationⁱ. Network slices are driven by service assurance requirements and are different from providing QoS to individual streams. Since the slice is end-to-end and touches almost all aspects of the network, many major cellular organizations are involved in its specifications.  In particular, ETSI is working on defining frameworks required for network slice implementations, GSMA is defining information models for slicing, and 3GPP is working on provisioning and resource management of RAN and core to support the slice. ETSI’s reference framework for a network slice is shown in Figure 1.

From the diagram, it is clear that at the highest levels, the network slices can dynamically come and go, at the middle layer, the VNFs can come, go or resize to support the traffic, while at the bottom layer, physical nodes and links can be added or fail;  with dynamic interactions among every level. A slice can have multiple alternate realization possibilities and during its life, the particular realization may have to change to another due to a variety of reasons, such as the come-up of another slice or the failure of a VNF or a physical node or link. To make the matter even more complex, it is expected that 5G deployments will include many small regional operators that provide the shared access network for the larger operators or enterprises operating their own 5G networks that might connect with bigger 5G networks for roaming or utilize the services of carrier-scale networks for failover and interconnectivity. For various business reasons, there is also industry momentum to move toward disaggregated nodes to allow for mixing and matching of network components.

 In the above dynamic environment, we critically need a mapping technique that can efficiently model the current state of the network and the relationships of a slice to VNF/software infrastructure/physical nodes. For the solution to scale, we need it to work across various domains and be dynamic and resilient in the event of a node failure or traffic surge. Further, it should be possible for a repository to add new nodes automatically and update the relationships to include new nodes of potentially different types. The repository should allow one to search based on desired properties or discover new relationships. Each node should be describable by its model and the repository must allow easy extensions to include new models across various operating domains.   

Traditionally, relational databases or tree-based schemas have been used to keep track of the relationships of services and devices in networks. The disadvantages of such schemas are that they cannot map and model the hierarchical relationship dynamically and cannot be used efficiently to explore indirect relationships that may exist between physical nodes and slices. Therefore, we need a new way to model the state of the network and its component relationships, within a layer or across layers.

Fortunately, there have been recent promising parallel developments in new schemas such as knowledge graphs in order to model, learn about, and investigate complex multi-relational settings. A knowledge graph describes objects of interest and connections between them while also providing a shared view of the knowledge within an organization. They are able to reuse definitions and descriptions that others create. Knowledge graphs are being used by companies such as Google, Facebook and LinkedIn to model and explore the complex relationships that could exist among words of a search query or entities in a social networkⁱⁱ. With physical networks reaching a similar level of dynamic behavior, complexity, and multi-layer-relationships, we should effectively evaluate the use of knowledge-graph based schemas to model the relationship between services, slices, VNFs and network nodes. These knowledge graphs should also take advantage of existing network semantics.

Having a knowledge graph-based repository will help us in modeling and exploring inter-layer and intra-layer relationships and discover new facts in network operations. These facts, available via simple queries or through automatic discovery, will assist human operators with root-cause analysis and planning, and will also help develop intelligent network applications and features to improve network efficiency. New networks are getting increasingly complex to be run by human operators and it is important to think of a unifying node and state repository that can match the complexity and dynamisms of today’s and tomorrow’s network.

References

[i] Next Generation Protocols (NGP) E2E Network Slicing Reference Framework and Information Model , ETSI GR NGP 011 v1.1.1 (2018-09) https://www.etsi.org/deliver/etsi_gr/NGP/001_099/011/01.01.01_60/gr_NGP011v010101p.pdf

[ii] Industry-Scale Knowledge Graphs: Lessons and Challenges By Natasha Noy, Yuqing Gao, Anshu Jain, Anant Narayanan, Alan Patterson, Jamie Taylor Communications of the ACM, August 2019, Vol. 62 No. 8, Pages 36-43