In the period 2014-2019 [1] the number of mobile-connected devices will grow from 7 Bn to 11 Bn, hence at a CAGR of 9%. However M2M/IOT  devices, which formed 6% of the 7 Bn devices in 2014, will constitute 28% of the 11 Bn devices in 2019... a market with much lower ARPU per connected device.  

In fact the number of smartphones and tablets will not exceed 5 Bn by 2019, and also there, a lower incremental ARPU is to be expected per device.  Cellular traffic will however continue to grow from 2.5 EB/month to 24 EB/month at 57% CAGR, implying that cellular devices will increasingly be used as router or relay for tethered Wi-Fi or Bluetooth devices.

Under these conditions Mobile Network Operators (MNOs) are thus facing 2 clear challenges : to reduce the cost per connected device and the cost per bit.  Since spectrum is a scarce and expensive resource, the latter depends a.o. on achieving greater spectral efficiency expressed in bit/s/Hz/cell.

The link with NFV and SDN appears when we consider that today’s radio and core networks are constituted of proprietary hardware, operating systems and software with limited lifetime and versatility.  It would probably be beneficial to MNOs to constitute pools of standardized resources (x86 computing, storage and networking) which can be shared among different network functions (such as virtualized RAN and EPC), service functions (such as a video optimizer), application functions (such as a usage monitoring app server) and IT systems (such as an OSS/BSS).  

There’s still some doubt whether NFV will initially achieve the necessary cost or power savings, but in the long run riding the x86 performance curve should arm MNOs in their competition with web-scale competitors.  And allow them to decentralize these functions, for low-latency access to decentralized content [2] or device-to-device communication.

At the same time it could be time for MNOs and the cellular industry in general to wonder about the root causes for slow growth in number of connected devices.  One factor hampering growth may be the SIM : 2G, 3G and 4G MNOs are unable to connect IOT devices or wearables, since the vast majority of these devices do not and will not contain any SIM.  If MNOs would permit access to their cellular networks via e.g. security certificates in the same way as they’re permitting employee access to their office Wi-Fi today, that barrier may be removed.

A second limiting factor is probably the connection-oriented mode.  Today’s number of mobile devices per cell is limited by the product of the consumed Data Radio Bearers (DRB) and bearer toggling rate, the latter being inversely proportional to the bearer inactivity timer.  The sooner an LTE eNodeB decides to liberate the DRB due to inactivity, the higher the toggling rate will be between ECM-active and ECM-idle mode, the more bearers will be available for use by other UE, but the more paging and other signalling traffic will be required to awaken the UEs from low-power state (ECM-idle) more frequently.  

As a result today in practice only 100-150 UE can be served per LTE cell, almost independently of the amount of spectrum. Serving more UE in connection-oriented mode would require more powerful RAN and EPC in both the user plane (number and depth of buffers...) and the control plane (signalling traffic).

3GPP is therefore partially abandoning the connection-oriented mode in its 3GPP Release 13, for the Cellular IOT [3], where DRB (and matching S1 bearers) will no longer be set up for CIOT UE, and low amounts of user data will instead be conveyed through the Signalling Radio Bearers (SRB), following successful authentication.  

The 3GPP work item and technology is referred to as NarrowBand IOT (NB-IOT) and in November 2015 a NB-IOT Forum was formed.

Before NB-IOT the 3GPP had already standardized multiple techniques helping MNOs to achieve a smooth transition to NFV and SDN :

1. Access Point Names (APNs) which the MNO can configure to be handled by a virtualized PDN Gateway (vPGW), e.g. a dedicated vPGW handling the IMS APN

2. GateWay Core Network (GWCN), which achieves the same effect by letting the eNodeB broadcast an additional Mobile Network Code (E.212 MNC), e.g. for IOT devices

3. Multiple Operator Core Network (MOCN), where based on the requested MNC the eNodeB can forward the attachment request to a dedicated (v)MME, (v)SGW and (v)PGW – i.e. a complete virtualized EPC (vEPC)

4. Dedicated Core Networks (DECOR), where the same effect can be achieved without broadcasting additional MNC – the initial MME analyzes the IMEI and redirects the attachment to a dedicated (v)EPC for a particular category of devices (IMEI range)

The better EPC vendors’ vPGW can partition 3GPP resources (number of PDN connections, bearers, throughput) among the APNs or Virtual APNs  used by multiple categories of subscribers or devices.  Within each APN one or several Policy Rule-Bases (PRB) contain Policy & Charging Control (PCC) rules or Application Detection & Control (ADC) rules specifying differentiated treatment (QoS) per Service and even per Application.  

In the better vendor’s vMME instance the paging policies can be configured per service category or service class (QCI).  The better RAN vendor supports midhaul between decentralized physical radio units processing the baseband, and virtualized RAN nodes (vRAN) handling PDCP.

Recently the Next-Gen Mobile Network (NGMN) Alliance [4], a group of RAN experts from MNOs, introduced the term “5G Slicing”, proposing in essence to spawn one 5G core network instance for each UE category.  It is unclear which 3GPP technique is proposed : APNs, GWCN, MOCN, DECOR, a combination, or a new one.

The fundamental issue with dedicated Virtualized Network Function Instances (VNFI) for different APNs, GWCN, MOCN, DECOR or “5G Slices” is definitely the orchestration and partitioning of computing, storage and networking resources among these VNFI.  

Indeed according to ETSI’s Management & Orchestration (MANO) architecture [5] the NFV Orchestrator (NFVO), VNF Manager (VNFM) and Virtual Infrastructure Manager (VIM) are unaware of the intra-VNF memory structures or short-term resource requirements.  Spawning a new VNFI, scaling it up, scaling it out or moving it to another compute node are supported but could take minutes, and by then the congestion may have been resolved.  

The granularity of such elasticity would also be too coarse to accommodate tens of “Slices” : a CPU core, a GB of RAM, a 10GE port...

As long as this issue remains unresolved we should rely on the VNF vendor to partition resources intelligently and fairly within each VNF Instance (VNFI), for user groups, services or applications.  And rely on the slower MANO elasticity process to expand/reduce a VNFI or spawn a new one.  A necessary step to achieve scale-in is for example to reduce the stateful information held within each VNFI, and store it in a central in-memory database – something which the NGMN Alliance is also proposing. 

The network vendor Huawei is however proposing to ignore the intra-VNFI resource partitioning challenge and to flee forward into a concept called Application-Dependent Networking (ADN) [6], in other words spawning  a set of dedicated VNF Instances for each Application accessed by the User Equipment (UE).  

In Huawei’s own words it is not more than a “vision”, but it is “relying on 5G network slicing”.  Whereas in NGMN’s definition each core network slice is reserved to a category of UE such as e.g. sensors, in Huawei’s intention a single UE would access multiple core network slices (sets of VNFI), one per Application.  An unmentioned shared node should then of course (based on deep packet inspection) steer downstream IP traffic from the Intenet via each core network slice to the UE...

In our opinion Huawei is confusing 5G Core Network Slicing with Service Chaining.  In the latter the UE’s context is dynamically created in a new 3GPP network element called the Traffic Detection Function (TDF) situated in the SGi LAN : the path between the core network (PDN Gateway) and the Packet Data Network (PDN i.e. the Internet).  

This TDF can analyze the traffic at L3, L4 and L7 and then steer a single UE’s upstream and downstream traffic through chains of non-virtualized and Virtualized Service Functions (VSF), one chain for each application.  Examples of VSF are virtualized firewalls, content caching or video optimization appliances.

We should collectively beware of vendors promising “Network Slicing” rather the 3GPP standard partitioning techniques.

Or promoting Network Slicing (separate VNFI) without even addressing the orchestration and resource partitioning issues between the slices; pushing these issues over the fence to the Mobile Network Operators.

Finally Huawei’s “Application Dependent Networking” (ADN) is a severely flawed and totally unjustified idea in which a core network slice is spawned per application, whereas it is in the SGi LAN that per-application Service Chaining would be required.  

Let’s instead start by addressing the intra-VNFI resource partitioning challenge, which has too often remained unimplemented, especially in Huawei’s core network products.

References

[1]    Cisco Visual Networking Index for Mobile, http://www.cisco.com/assets/sol/sp/vni/forecast_highlights_mobile/index.html
[2]    Korea Telecom’s 5G strategy, https://www.netmanias.com/en/post/oneshot/8185/5g-c-ran-kt-korea/kt-5g-network-architecture
[3]    3GPP TR 23.720, TR 45.820
[4]    NGMN Alliance 5G White Paper section 5.4, https://www.ngmn.org/uploads/media/NGMN_5G_White_Paper_V1_0.pdf
[5]    ETSI Management & Orchestration architecture, https://www.sdxcentral.com/resources/nfv/nfv-mano/
[6]    Huawei ADN, http://www.huawei.com/en/news/2015/12/Huawei Unveils the Vision for