Rakuten makes a real racket about its virtual network at MWC

Open RAN went commercial in intent and deployment at MWC 2019. TMN is taking a multi-part look at the open RAN news, actualisation and commercialisation on show at MWC Barcelona 2019.

Open RAN Part II: Ra Ra Rakuten 

One operator that has recently snuck up on the rails for many outside of Japan is Rakuten. But there was no missing the new entrant just prior to and at MWC as it featured in a number of conference speaking opportunities and hosted and took part in a  further number of side events. The mission was to shout loudly about the potential the industry has to transform how it builds and operates mobile networks. 

In the weeks preceding MWC, TMN picked up on the extent of the company’s radical vendor selection policy in the RAN. Rakuten decided last year to go for an all virtual RAN, on Intel-based hardware, and not to do that via any traditional RAN supplier. Instead it is putting the first virtual Distributed Units into 4,000 edge sites,  with all the vRAN software from Altiostar. To back that up, it has taken a strategic investment in Altiostar, making it doubly in its interest that Altiostar’s technology is a success.

Nokia is supplying the remote radio units, bolting these on via a specially designed coupling to antennas from Korean antenna supplier KMW Communications. Cisco is the cloud core and NFVi provider. The hardware, which comes in just four server variants, is all from Taiwanese provider Quanta Cloud Technology (as first mentioned in TMN at last year’s MWC delivering an Intel-based next gen Central Office solution).

Red Hat is providing the Open Stack based cloud operations and automation control. Intel seems to have worked its tail off to make sure it can support the workloads Rakuten is placing in its hands. (More on this to follow)

Although the company is not explicitly building its RAN to formal TIP or O-RAN specifications, its insistence is that any software must be able deployable in its cloud environment, and manageable by its cloud management and automation software. It also has different vendors from core, to the RAN processing to the radiohead and antennas.

Most notably, thrusting CTO Tareq Amin was suddenly everywhere, providing the “networks” bit of MWC with its noisiest story. That’s partly a function of the number of vendor engagements the company has (is he making part-payment to his vendors in PR?) So he was at Cisco’s soiree on Sunday, telling media how he has revolutionised the capex and operational paradigm (35% less TCO than even the next best network) by building the first all virtual cloud telco network for no money, with 100 operations staff.

He was on a CTO panel in the main Conference on Monday, telling delegates how he has revolutionised the capex and operational paradigm by building the first all virtual cloud telco network for not much money, with 100 operations staff.

He was on the operator’s bright white and red booth later that day, showing Japanese media a demonstration of how… you get the gist. He was back on stage on Rakuten’s booth on Wednesday lunchtime with five of his vendor partners, delivering the same message.

It must be a bit galling for the existing operators to see this man, who was also a large part of the last next best thing – Reliance Jio in India – enter the scene and start taking aim at their outmoded and outdated ways of thinking.

And those thoughts in summary: this industry is afraid of innovation, it is terrified of and terrible at supporting new startups, only humans get in the way of change, being a brownfield operator with legacy to transform is just an excuse, you could change if you really wanted to.

To be fair to Amin, there’s a lot more to his presence and drive than self-promotion – it seems that there is a genuine desire on behalf of his vendor partners to push the industry to consider the disaggregated, open hardware and software approach at scale and for real – not as lip service and not as a science project.

He knows much of what Rakuten has achieved is about mindset because he has already had to change mindsets, he said, or to find those who aligned with his.

On Sunday he said he knew Cisco wasn’t technically ready to support and roll out a totally virtualised core network on NFVi across two large core datacentres and 4,000 distributed sites, but he went with the vendor because he knew it would adopt the correct mindset. On his CTO panel, he said, “In RAN, if I went to the traditional large OEMs they would say, ‘You are crazy. You cannot run a vCU on x86 – it will not work’.

“So we said we would think differently and partnered and took a big position in Altiostar, believing they have the right software mindset. And we built the world first vDU – so the actual site is down to very little. There’s no cabinet or big equipment or baseband. The site is a very simplified antenna and the Radio Unit, with dark fibre connecting it to the edge. It works and it works remarkably well.

“The capex impact is huge, but the operational efficiency is even more impactful. Everything is automated. From taking eight hours to activate one traditional NodeB, when the RAN is virtualised it takes the activation of a VM to 15 minutes.”

Tremendous fun, anyway. And more to come. But questions will arise. One will relate to whether this deployment can truly be replicable in other markets. Will Altiostar’s Option 7 split, which requires a lot of dark fibre fronthaul capacity, work for it in other markets? And can operators really replicate Rakuten’s deployment in terms of the assets it has available? After all, is it not a bit of a one-off with its ability to build a greenfield network, with its large number of existing points of presence in which to place its RAN compute and edge platforms, and its access to dense fibre assets with which to connect those edge PoPs to the Radio Units and its two large core datacentres? It also has a large IT operation that already understood how to run things like microservices and containerisation (according to Amin at any rate).

Here’s Amin answering some of these points directly, and also giving some more detail on the rollout.

Is the rollout really replicable for legacy operators, and just about mindset?
I don’t think there is any technical challenges at all. It does not mean that it is trivial or easy to do but what I do believe is if you start looking at workloads as software it becomes easier for us to identify problems, isolate, resolve and fix in an agile manner. If I did not have the right support from my organisation none of this stuff here you would see. We started in August last year and here in February we are almost 80% through the major software issues we have to address. So I don’t think it’s technically that complex.

“The large OEMs have a bigger problem. They have to contemplate how they will transform the organisation. I think that’s a much bigger challenge structurally than technological issues.”

Can you really run your workloads straight on Intel based hardware?
“It’s not trivial because radio access, L1, requires ultra low latency to manage the radio access signalling. The challenges to do so requires you to harden the Open Stack NFVi layer. Intel had to do a lot of work on the DPDK layer, and of course enable hardware acceleration through their FPGAs.”

What is the capex budget for the rollout?
“We talk about $5-6 billion range to build a network with the same size infrastructure as Sprint in the US. But the upgrade to 5G will be 60% cheaper because of this architecture.”

How do the 4,000 edge nodes map to sector coverage and the number of radio heads you are deploying?
“Traditional infrastructure has a baseband with one to one mapping which is highly inefficient. We have eight sites mapped to a single server but it does not mean a server could only serve these eight sites. The way we did the architecture this NFVi layer is horizontal. So think of the radio as a VNF, we can scale out as many servers as you want – but they are multi-purpose and distributed so you are able to leverage multiple sites, not just one to one mapping.

“We need to divorce the mentality that there is the BBU, the BBU becomes software and whatever is running this hardware the software uses the server it needs. So I cannot say it is one server to one site or one to many – it’s just depending on the workload it gets allocated the compute or the hardware accelerator functions that you require.

How many radio heads will you install?
“If you look at the scale, roughly speaking 37,000 eNodeB times three [111,000] is the amount of radio heads we will have just for LTE only. It’s very large.”

One of the key partners, Cisco, was also highly challenged to provide that automated, cloud-based architecture. Cisco’s Bob Everson –  Global Director of Mobility and 5G, told TMN how he thinks the industry might apply the lessons it has learnt with Rakuten.

“When you look at the architecture it’s the first proof at scale that you can do this in a different way. It’s a fully virtualised, cloud based architecture, with all the RAN elements moved into software and the whole thing orchestrated as one software based system. It’s shifting people’s mental model, I think, of the way a mobile network should be built.

You have to get beyond the cynical “Oh it must be nice for you” to, “Look at all the pieces in here that I can replicate

“Nobody’s ever done this before. We are doing things that are unprecedented. Just virtualising the DU, the lower layers of the radio stack: a few months ago there were vendors that said we would never be able to do it. But we are and we have learned a lot along the way: as we go forward we are figuring out how to scale it and build practices around it.

“Yes, other operators can replicate that. We get this question a lot from operators as well. The thing with Rakuten is it is a whole system with a bunch of different elements that are replicable. Sure there are very few scenarios that look exactly like theirs but if you look at the platform they have built out, the programmable infrastructure piece can be reused.

“Is it going to be day one across an existing legacy macro network? No. But a lot of these customers are saying, ‘OK as I move towards 5G, maybe I will start with mmWave, or midband spectrum, or do small cells in this area.’ So if you pick off one of those use cases to start with you can build that as an overlay. In some cases customers are saying they will do this in 5G SA; use automation there and then start to bring over the legacy infrastructure over time as we prove that out. It’s about taking the pieces and working out where the right insertion point is. But you have to get beyond the cynical “Oh it must be nice for you” to, “Look at all the pieces in here that I can replicate”.

“The functional split they are using, Option 7.2 that is also now in O-RAN, seems to be the one people are consolidating around. I know one of the major T1 operators just told all their legacy vendors they were going to require support for that from them –  that everything, even if it’s just from the base of tower to top of tower, had to be based on that split. But even if that changes, it’s much easier to change an interface if you disaggregate that from the start.”