The xRAN Forum is a carrier-led initiative aiming to apply the principles of virtualization, openness and standardization to one area of networking that has remained stubbornly closed and proprietary -- the radio access network (RAN) and, in particular, the critical segment that connects a base station unit to the antennas. Recently, I sat down with Dr. Sachin Katti, Professor in the Electrical Engineering and Computer Science departments at Stanford University and Director of the xRAN Forum, to find out what this is all about.
Jim Carroll, OND: Welcome Professor Katti. So let's talk about xRAN. It's a new initiative. Could you introduce it for us?
Dr. Sachin Katti, Director of xRAN Forum: Sure. xRAN is a little less than two years old. It was founded in late 2016 by me along with AT&T, Deutsche Telecom and SK Telecom -- and it's grown significantly since then. We now are up to around ten operators and at least 20 vendor companies so it's been growing quite a bit the last year and a half.
JC: So why did xRAN come about?
SK: Some history about how all of happened... I was actually at Stanford as my role as a faculty here at Stanford collaborating with both AT&T and Deutsche Telecom on something we called soft-RAN, which stood for software-defined radio access network. The research really was around how do you take radio access networks, which historically have been very tightly integrated and coupled with hardware, and make them more virtualized - to disaggregate the infrastructure so that you have more modular components, and also defined interfaces between the different common components. I think we all realized at that point that to really have an impact, we need to take this out of the research lab and get the industry and the cross-industry ecosystem to join forces and make this happen in reality.
That's the context behind how xRAN was born. The focus is on how do we define a disaggregated architecture for the RAN. Specifically, how do you take what's called the eNodeB base station and deconstruct the software stuff that's running on the base station such that you have modular components with open interfaces between them that allows for interoperability, so that you could truly have a multi-vendor deployment. And two, it also has a lot more programmability so that an operator could customize it for their own needs, enabling new applications and new service much more easily without having to go through a vendor every single time. I think it was really meant so that you can try all of those aspects and that's how it got started.
JC: Okay. Is there a short mission statement?
SK: Sure. The mission statement for xRAN is to build an open virtualized, disaggregated radio access network architecture that opens standardized interfaces between all of these components, and to be able to build all of these components in a virtualized fashion on commodity hardware wherever possible.
JC: In terms of the use cases, why would carriers need to virtualize their RAN, especially when they have other network slicing paradigms under development?
SK: It's great that you bring up network slice actually. Network slicing is one of the trialing use cases and the way to think about this is, in the future, everyone expects to have network slices with very different connectivity needs for enabling different kinds of applications. So you might have a slice for cars that have very different bandwidth and latency characteristics compared to a slice for IOT traffic, which is a bit more delay tolerant for example.
JC: And those are slices in a virtual EPC? Is that right?
SK: Those are slices that need to be end-to-end. It can't just be the EPC because ultimately the SLAs you can give for the kind of connectivity you can deliver, is ultimately going to be dictated by what happens on the access. So, eventually, a slice has to be end-to-end and the challenge was if an operator, for example, wants to define new slices then how do they program the radio access network to deliver that SLA, to deliver that connectivity that that slice needs.
In the EPC there was a lot of progress on what are those interfaces to enable such slicing but there was not similar progress that happened in the RAN. How do you program the base station, and how do you program the access network itself to deliver such slicing capability? So that's actually one of the driving use cases that's in there since the start of xRAN. Another big use case, and I'm not sure whether we should call it a use case, but just a need, is around having a multi-vendor deployment. Historically, if you look at radio access network deployments, they're a single vendor. So, if you take a U.S. operator, for example, they literally divide up their markets into an Ericsson market or a Nokia market or whatever. And the understanding is everything in that market, from the base station to the antenna to the backhaul, everything comes from one vendor. They really cannot mix and match components from different vendors because there haven't been many interoperable interfaces, so the other big need or requirement that is coming all this is interoperability in a multivendor environment that they want to get to.
JC: How about infrastructure sharing? I mean we see that the tower companies are now growing by leaps and bounds and many carriers thinking that maybe it's no longer strategically important to own the tower and so share that tower, and they might share the backhaul as well.
SK: It will actually help. It will actually enable that kind of sharing at an even more deeper level, because if you have an infrastructure that is virtualized and is running on more commodity hardware in a virtualized fashion then it becomes easier for a tower company to set up the compute substrate and their underlying backhaul substrate and then provide virtual infrastructure slices to each operator to operate on top of. And so instead of actually just physically separating -- right now they are basically renting space on the top right but instead if you could just the same underlying compute substrate and the same backhaul infrastructure as well a fronthaul infrastructure and virtually slice it and run multiple networks on top, it actually makes it possible to share on the infrastructure even more. So virtualization is almost a prerequisite to any of the sharing of infrastructure.
JC: Tell us about the newly released, xRAN fronthaul specification version 1.0. What is the body of work it builds on?
SK: Sure, let me step back and just talk about all the standardization efforts, and then I'll answer the question. xRAN actually has three big different working groups. One is around fronthaul, which refers to the link between the radio head and that baseband unit. This is the transport that's actually carrying the data between the baseline unit and the radio transmission and, in the reverse direction, when you receive something from the mobile unit. So that's one aspect. The second one is around the control plane and user plane separation in the base station. Historically, the control plane and the user plane are tightly coupled. A significant working group effort in xRAN right now is how do you decouple those and define standardized interfaces between a control plane and a user plane. And the last working group is trying to define what are the interfaces between the control plane of the radio access network and orchestration systems like ONAP. So those are three main focus areas.
Our first specification, which describes the fronthaul interfaces, was released this month. So, what went on there? The problem that we solved concerns closed interfaces. Today if you bought a base station you also have to buy the antenna from the same vendor. That's it. For example, if you bought an Ericsson base station you have to buy an antenna from Ericsson as well. There are very few compatible antenna systems, but with 5G, and even with 4G, there's been a lot of innovation on the antenna side. There are innovators developing massive MIMO systems. These have lots of antennas and can significantly increase the capacity of the RAN. Many start-ups that are trying to do this, but they're struggling to get any traction because they cannot sell their antennas and connect it to an existing vendor's baseband unit. So, a critical requirement that operators are pushing was how do we make it such that this fronthaul specification is truly interoperable, making it possible to mix and match. You could take a small vendor's radio head and antenna and connect it with an existing well-established vendor's baseband unit -- that was the underlying requirement. What the new fronthaul work is truly trying to accomplish is to make sure that this interface is very clearly specified such that you do not need tight integration between the baseband unit and the radio head unit.
This fronthaul work came about initially with Verizon, AT&T and Deutsche Telekom driving it. Over the past year, we have had multiple operators joining the initiative, including NTT DoCoMo, and several vendors they brought along including Nokia. Samsung, Mavenir, and a bunch of other companies, all coming together to write the specification and contribute IP towards it.
JC: Interesting, so you have support from those existing vendors who would seem to have a lot to lose if this disaggregation occurred disfavorably to them.
SK: Yes, we do. Current xRAN members include all or the bigger vendors, such as Nokia and Samsung, especially on the radio side. Cisco is a member which is more often on the orchestration side and there are several other big vendors that are part of this effort. And yeah, they have been quite supportive.
The xRAN Forum is an operator-driven body. The way we set up a new working group or project is that operators come in and tell us what their needs are, what their use cases are, and if we see enough consistency, when multiple operators share the same need or share the same use case, that leads to the start of the new working group. The operators often end up bringing their vendors along by saying we need this, "we are gonna drive it through the xRAN consortium and we need you to come and participate, otherwise you'll be left out." That's typically how vendors are forced to open up.
JC: Okay, interesting, so let's talk a little bit about the timelines and how this could play out. You talked about plugging into an existing baseband unit or base station unit so I guess there is a backward compatibility aspect?
SK: No, we are not expecting operators to build entirely new networks. The first fronthaul specification is meant both for 4G and 5G. The fronthaul is actually independent of the underlying air interface so it can work under 4G networks. On the baseband side, it does require a software update. It does require these systems to adhere to the spec in terms of how to talk to the radio head, and if they do, then the expectation is that someone should be able to plug in a new radio head and be able to make that system work. That being said, where we are at right now, is we have released a public specification. We believe it's interoperable but the next stage is to do interoperability testing. We expect that to happen later this year. Once interoperability testing happens, we will know what set of systems are compatible. Then we will have, if you will, a certificate saying that these are compliant.
JC: And would that certification be just for the fronthaul component or would that be for the control plane and data plane separation as well?
SK: Our working groups are progressing at different cadences. The fronthaul specification already is out and they expect to the interoperability testing later this year, and that will be only for the fronthaul. As and when we release the first specification for the control plane and use plane separation, we will have a corresponding timeline. But I think one thing to realize is that these are not all coupled. You could use the fronthaul specification on its own without having the rest the architecture. You could take existing infrastructure implement just the fronthaul specification and realize the benefits of the interoperability without necessarily having a control plane that's decoupled from the user plane. So the thing is structured such that each of those working groups can act independently. We didn't want to couple them because that would mean that it'll take a long time before anything happens.
JC: Wouldn't some of the xRAN work naturally have fit into 3GPP or ETSI's carrier virtualization efforts? Why have a new forum?
SK: Definitely. 3GPP is a big intersection point. I think the way we look at it is that we are trying to work on areas that 3GPP elected not to. So if it has anything to do with the air interface, for example, how should the infrastructure talk to the phone itself -we are not trying to work in that space. If it's got anything to do with how the base station talks to the core network, we are not trying to specify that interface. But there are things that 3GPP elected not to work on for whatever reason, and which could be how vendor incentives come into play. Perhaps these vendors discouraged 3GPP from working on intereroperable fronthaul interfaces. And we don't know the reason why 3GPP chose this path. You can see that this is also operator driven. So operators want certain things to happen but they
are not successful in getting 3GPP to do it. So xRAN is a venue for them to come in and specify
what they want to do and what they want to accomplish and get appropriately incentivized
vendors to actually come up together. So it is complementary in terms of the work effort, but I could see a scenario where the fronthaul specification that we come out with, this one and the next one, eventually forms the basis for a 3GPP standardized specification -- but that's not necessarily a conflict -- that actually might be how things eventually get fully standardized.
JC: There are other virtualization ideas that have sprung up from this same lab and in the Bay Area. How does this work in collaboration with CORD and M-CORD?
SK: Historically, I think virtualization has infected, if you will, the rest of the networking domain but has struggled to make headway in the RAN. If you looked at the rest of the network there's been a lot of success with virtualization. The RAN has traditionally been quite hard to do. I think there are multiple reasons for that. One is that the workload -- the things that you want to do in their RAN -- are much more stressful and demanding than the rest of the network in terms of processing. I think the hardware is now catching up to the point where you can take off-the-shelf hardware and run virtualized instances of the RAN on top. I think that's been one.
Second, the RAN is also a little bit harder to disaggregate because many of the control plane
decisions are occurring at a very fast timescale. There are things, for example, like how should I
schedule a particular user’s traffic to be sent over the air. That's a decision that the base station is making every millisecond and, at that timescale, it's really hard to run it at a deeper level. So, having a separate piece of logic making that decision, and then communicating that decision to the data plane if you will, and then the data plane implementing that decision, which would be classically how we think about SDN, that's not going to work because if you have a round-trip latency of one millisecond that you can tolerate, it's too stringent. I think we need to figure out how to deconstruct the problem, take out the right amount of control logic but still leave the very latency sensitive pieces in the underlying data plane of the infrastructure itself. I think that's still work in progress. We still know there are hard technical challenges there.
JC: Okay, talking about inspiration -- one last thing- is there an application that you have in
mind that inspires this work?
SK: Sure. I am thinking a pretty compelling example is network slicing. As you look at these very demanding applications --if you think about virtual reality and augmented reality applications, or self-driving cars --there are very strict requirements on how that traffic should be handled in the network. If I think about a self-driving car, and it wants to offload some of its some mapping and sensing capabilities to the edge cloud, that loop, that interaction loop between that car and the edge cloud has very strict requirements. And you want that application to be able to come to the network and say this is the kind of connectivity I need for my traffic, and for the network to be programmable enough that the operator should be able to program the underlying infrastructure such that I can deliver that kind of connectivity to the self-driving car application.
I think those two classes of applications are characterized by latency sensitivity and bandwidth intensity. You don't get any leeway on either dimension. Right now, the people developing those applications do not trust the network. If you think about current prototypes of self-driving cars, the developers cannot assume that the network will be there. So they currently must build very complex systems to make the vehicle completely autonomous. If we truly want to build thinks where the cloud can actually play a role in controlling some systems, then we need this programmable network to enable such a world.
Excellent, well thank you very much and good luck!
