BT's McRae: Common Service Models Aid Automation

BT is relying on common service models to provision new dynamic services while also bringing cohesion to its NFV and SDN efforts.

Service models, which are based on TOSCA and YANG languages, provide carriers with a common context for building solutions, according to BT's Neil McRae, chief network architect. McRae said that one of the reasons that BT has been successful with automation is due to having a foundation in place with service models.

In addition to building new services and products today, McRae said that common service models would play a key role in enabling artificial intelligence and machine learning in the future.

In Part II of this Q&A series, McRae also spoke about how his company is currently implementing machine learning. In Part I, he discussed how SDN and automation were working together to make networks and services more agile. (See BT's McRae: SDN Is All About Automation .)

Telco Transformation: Can you describe BT's use of common service models, and is BT using them more than others in the industry?

Neil McRae: We maybe focus on this as more of a critical requirement than others do. I think the industry is a bit confused right now around NFV, SDN and other software-based technologies. And many people are trying to do things, and find it difficult. I believe part of that goes to the starting point of what are we trying to enable through this softwarization of network services. Then you add things like white boxes, and programmable silicon, and very quickly you find yourself in potentially a very complex environment.

But I think, in my mind, I've always been: When I push something, I want to be really clear about where I ought to be in this software world. How do I enable this system of networks to work in that software world? We're not the only ones that think like that. Some of the big web pseudowire guys also believe this. There's been Google doing a lot of work in this space, as well.

We believe that the service models are crucial now. We've kind of been working on that for a while, and at a various pace. We could say we've been flat out, that it's the only thing we've been doing, but it's been a critical thing for us in some of the new dynamic network services products we've been working on.

TT: What are some examples of those new dynamic services?

NM: So our SD-WAN product is in trial with some customers now. Really, for us, the way we've been able to really make a success of the automation of that is by having those service models well-defined.

Now I think the industry's going towards SDN and NFV in generic terms. They're both highly valuable technologies, but I think our approach into them has been perhaps not as optimized as it could have been. We believe that it's really down to the fact that if you're going to virtualize something -- and I learned this the hard way when we built data centers for virtualization for IT -- you need to be really clear about what you're doing. The non-virtualized world that you are in to the virtualized world that you are going to move to, you need to be really clear about what's in that and what the shape of that is. That's where, I think, the service models can help us.

They don't solve every single problem, but I think they give you a common context. If you're working with a vendor, or a developer, or a network engineer, they give you a common context that you could start to build solutions to. I believe that having those models will be critical for that machine learning, AI future that we're all very excited about. I think that trying to have automation work on something that isn't well-defined means that you have automation that is very disport and very expensive to manage.

TT: How do you build automation and machine learning into your network today?

NM: We're using machine learning in a couple of ways; both in the network, but also more generically in our IT systems. This starts with the service model piece.

If you look at car manufacturers, what are they good at? They're great at building automated plants that build cars. They're phenomenal at it, right? How did they do that? With a model of the car. That's how they do it. From that model, they can automate, use machine learning, robotics, and a bunch of other things. This is kind of the genesis of the service models for me, which is, if we could build a model of our services then I could use that to teach things on machine learning and automation.

And for that, my network operations teams can be the teachers in this world. They will teach the automation software about the models that we built of the service. So, take Ethernet -- a very simple service-- it's a port on a router at the A end, it's a port on the router at the B end, and a pseudowire. That's a relatively simple model, and it can be a very simple concept for machine learning to accomplish.

Also, it can be used as a model for a basis of comparison. If there's something not right with the service, it can use that model to say, "Okay, this piece here, the pseudowire, isn't working. Why is that?" It's because actually there's a circuit down on our backboard network, which reroutes it. That would happen very quickly, and that's a very simple approach to this.

So, if we think of VPN networks, cloud business, and solutions being built into those VPN networks, then security on-call, firewalls, queuing protection, etc., it starts to get very complex. That's where, I believe if you've got a very tight set of models of the services, you can use that automation in a way that we've tried to do in the past, but found it very difficult because our starting point for the service has never been that clear. I think, fundamentally, in the future service models replace what we know of the service industry.

— Mike Robuck, Editor, Telco Transformation