Recognized now as a cognitive bias known as the frequency illusion, this phenomenon is thought to be evidence of the brain’s powerful pattern-matching engine in action, subconsciously promoting information you’ve previous deemed interesting or important. While there is far from anything powerful between my ears, I think my brain was actually on to something. As the need to support an increasingly diverse array of equally critical but diverse services and endpoints emerges from the 4G ashes, network slicing is looking to be a critical function of 5G design and evolution.
Euphoria subsiding, I started digging a little further into this topic and it was immediately apparent that the source of my little bout of déjà vu could stem from the fact that network slicing is in fact not one thing but a combination of mostly well-known technologies and techniques… all bundled up into a cool, marketing-friendly name with a delicately piped mound of frosting and a cherry on top. VLAN, SDN, NFV, SFC — that’s all the high-level corporate fluff pieces focused on. We’ve been there and done that.2
An example of a diagram seen in high-level network slicing fluff pieces
I was about to pack up my keyboard and go home when I remembered that my interest had originally been piqued by the prospect of researching RAN virtualization techniques, which must still be a critical part of an end-to-end (E2E) 5G network slicing proposition, right? More importantly, I would also have to find a new topic to write about. I dug deeper.
A piece of cake
Although no one is more surprised than me that it took this long for me to associate this topic with cake, it makes a point that the concept of network slicing is a simple one. Moreover, when I thought about the next step in network evolution that slicing represents, I was immediately drawn to the Battenberg. While those outside of England will be lost with this reference,3 those who have recently binge-watched The Crown on Netflix will remember the references to the Mountbattens, which this dessert honors.4 I call it the Battenberg Network Architecture Evolution principle, confident in the knowledge that I will be the only one who ever does.
The Battenberg Network Architecture Evolution Principle™
Network slicing represents a significant evolution in communications architectures, where totally diverse service offerings and service providers with completely disparate traffic engineering and capacity demands can share common end-to-end (E2E) infrastructure resources. This doesn’t mean simply isolating traffic flows in VLANs with unique QoS attributes; it means partitioning physical and not-so-physical RF and network functions while leveraging microservices to provision an exclusive E2E implementation for each unique application.
Well, consider the Internet of Talk vs. the Internet of Things, as the subtitle of the post intimates. Evolving packet-based mobile voice infrastructures (i.e. VoLTE) and IoT endpoints with machine-to-person (M2P) or person-to-person (P2P) communications both demand almost identical radio access networks (RAN), evolved packet cores (EPC) and IP multimedia subsystem (IMS) infrastructures, but have traffic engineering and usage dynamics that would differ widely. VoLTE requires the type of capacity planning telephone engineers likely perform in their sleep, while an IoT communications application supporting automatic crash response services5 would demand only minimal call capacity with absolutely no Mother’s Day madness but a call completion guarantee that is second to none.
In the case of a network function close to my heart — the IMS Core — I would not want to employ the same instance to support both applications, but I would want to leverage a common IMS implementation. In this case, it’s network functions virtualization (NFV) to the rescue, with its high degree of automation and dynamic orchestration simplifying the deployment of these two distinct infrastructures while delivering the required capacity on demand. Make it a cloud-native IMS core platform built on a reusable microservices philosophy that favors operating-system-level virtualization using lightweight containers (LCXs) over virtualized hardware (VMs), and you can obtain a degree of flexibility and cost-effectiveness that overshadows plain old NFV.
I know I’m covering a well-trodden trail when I’m able to rattle off a marketing-esque blurb like that while on autopilot and in a semi-conscious state. While NFV is a critical component of E2E network slicing, things get interesting (for me, at least) when we start to look at the virtualization of radio resources required to abstract and isolate the otherwise common wireless environment between service providers and applications. To those indoctrinated in the art of Layer 1-3 VPNs, this would seem easy enough, but on top of the issue of resource allocation, there are some inherent complications that result from not only the underlying demand of mobility but the broadcast nature of radio communications and the statistically random fluctuations in quality across the individual wireless channels. While history has taught us that fixed bandwidth is not fungible,6 mobility adds a whole new level of unpredictability.
The Business of WNV
Like most things in this business, the division of ownership and utilization can range from strikingly simple to ridiculously convoluted. At one end of the scale, a mobile network operator (MNO) partitions its network resources — including the spectrum, RAN, backhaul, transmission and core network — to one or more service providers (SPs) who use this leased infrastructure to offer end-to-end services to their subscribers. While this is the straightforward MNV model and it can fundamentally help increase utilization of the MNOs infrastructure, the reality is even easier, in that the MNO and SP will likely be the same corporate entity. Employing NFV concepts, operators are virtualizing their network functions to reduce costs, alleviate stranded capacity and increase flexibility. Extending these concepts, isolating otherwise diverse traffic types with end-to-end wireless network virtualization, allows for better bin packing (yay – bin packing!) and even enables the implementation of distinct proof-of-concept sandboxes in which to test new applications in a live environment without affecting commercial service.
Breaking down the 1-2 and 4-layer wireless network virtualization business model
Continuing to ignore the (staggering, let us not forget) technical complexities of WNV for a moment, while the 1-2 layer business model appears to be straightforward enough, to those hell-bent on openness and micro business models, it appears only to be monolithic and monopolistic. Now, of course, all elements can be federated.7 This extends a network slice outside the local service area by way of roaming agreements with other network operators, capable of delivering the same isolated service guarantees while ideally exposing some degree of manageability.
To further appease those individuals, however, (and you know who you are) we can decompose the model to four distinct entities. An infrastructure provider (InP) owns the physical resources and possibly the spectrum which the mobile virtual network provider then leases on request. If the MVNP owns spectrum, then that component need not be included in the resource transaction. A widely recognized entity, the mobile virtual network operator (MVNO) operates and assigns the virtual resources to the SP. In newer XaaS models, the MVNO could include the MVNP, which provides a network-as-a-service (NaaS) by leveraging the InPs infrastructure-as-a-service (IaaS). While the complexities around orchestration between these independent entities and their highly decomposed network elements could leave the industry making an aaS of itself, it does inherently streamline the individual roles and potentially open up new commercial opportunities.
Dicing with RF
Reinforcing a long-felt belief that nothing is ever entirely new, long before prepending to cover all things E2E, the origin of the term “slicing” can be traced back over a decade in texts that describe radio resource sharing. Modern converged mobile infrastructures employ multiple Radio Access Technologies (RATs), both licensed spectrum and unlicensed access for offloading and roaming, so network slicing must incorporate techniques for partitioning not only 3GPP LTE but also IEEE Wi-Fi and WiMAX. This is problematic in that these RATs are not only incompatible but also provide disparate isolation levels — the minimum resource units that can be used to carve out the air interface while providing effective isolation between service providers. There are many ways to skin (or slice) each cat, resulting in numerous proposals for resource allocation and isolation mechanisms in each RF category, with no clear leaders.
At this point, I’m understanding why many are simply producing the aforementioned puff pieces on this topic — indeed, part of me now wishes I’d bowed out of this blog post at the references to sponge cake — but we can rein things in a little. Most 802.11 Wi-Fi slicing proposals suggest extending existing QoS methods — specifically, enhanced DCF (distributed coordination function) channel access (EDCA) parameters. (Sweet! Nested acronyms. Network slicing might redeem itself, after all.) While (again) not exactly a new concept, the proposals advocate implementing a three-level (dimensional) mathematical probability model know as a Markov chain to optimize the network by dynamically tuning the EDCA contention window (CW), arbitration inter-frame space (AIFS) and transmit opportunity (TXOP) parameters,8 thereby creating a number of independent prioritization queues — one for each “slice.” Early studies have already shown that this method can control RF resource allocation and maintain isolation even as signal quality degrades or suffers interference. That’s important because, as we discussed previously, we must overcome the variations in signal-to-noise ratios (SNRs) in order to effectively slice radio frequencies.
In cellular networks, most slicing proposals are based on scheduling (physical) resource blocks (P/RBs), the smallest unit the LTE MAC layer can allocate, on the downlink to ensure partitioning of the available spectrum or time slots.
An LTE Physical Resource Block (PRB), comprising 12 subcarriers and 7 OFDM symbols
Slicing LTE spectrum, in this manner, starts and pretty much ends with the eNodeB. To anyone familiar with NFV (which would include all you avid followers of Metaswitch), that would first require virtualization of that element using the same fundamental techniques we’ve described in numerous posts and papers. At the heart of any eNodeB virtualization proposition is an LTE hypervisor. In the same way classic virtual machine managers partition common compute resources, such as CPU cycles, memory and I/O, an LTE hypervisor is responsible for scheduling the physical radio resources, namely the LTE resource blocks. Only then can the wireless spectrum be effectively sliced between independent veNodeB’s owned, managed or supported by the individual service provider or MVNO.
Virtualization of the eNodeB with PRB-aware hypervisor
Managing the underlying PRBs, an LTE hypervisor gathers information from the guest eNodeB functions, such as traffic loads, channel state and priority requirements, along with the contract demands of each SP or MVNO in order to effectively slice the spectrum. Those contracts could define fixed or dynamic (maximum) bandwidth guarantees along with QoS metrics like best effort (BE), either with or without minimum guarantees. With the dynamic nature of radio infrastructures, the role of the LTE hypervisor is different from a classic virtual machine manager, which only need handle physical resources that are not continuously changing. The LTE hypervisor must constantly perform efficient resource allocation in real time through the application of an algorithm that services those pre-defined contracts as RF SNR, attenuation and usage patterns fluctuate. Early research suggests that an adaptation of the Karnaugh-map (K-map) algorithm, introduced in 1953, is best suited for this purpose.9
Managing the distribution of these contracted policies across a global mobile infrastructure falls on the shoulders of a new wireless network controller. Employing reasonably well-understood SDN techniques, this centralized element represents the brains of our virtualized mobile network, providing a common control point for pushing and managing policies across highly distributed 5G slices. The sort of brains that are not prone to the kind of cognitive tomfoolery that plague ours. Have you ever heard of the Baader-Meinhof phenomenon?
1. No one actually knows why the phenomenon was named after a West German left wing militant group, more commonly known as the Red Army Faction.
3. Quite frankly, as a 25-year expat and not having seen one in that time, I’m not sure how I was able to recall the Battenberg for this analogy.
4. Technically, it’s reported to honor of the marriage of Princess Victoria, a granddaughter of Queen Victoria, to Prince Louis of Battenberg in 1884. And yes, there are now two footnotes about this cake reference.
5. Mandated by local government legislation, such as the European eCall mandate, as I’ve detailed in previous posts. https://www.metaswitch.com/the-switch/guaranteeing-qos-for-the-iot-with-the-obligatory-pokemon-go-references
6. E.g. Enron, et al, and the (pre-crash) bandwidth brokering propositions of the late 1990s / early 2000s
7. Yes — Federation is the new fancy word for a spit and a handshake.
8. OK – I’m officially fully back on the network slicing bandwagon.
9. A Dynamic Embedding Algorithm for Wireless Network Virtualization. May 2015. Jonathan van de Betl, et al.