Between two and three hundred people attended the ETSI Network Functions Virtualization (NFV) meeting in Santa Clara last month. As expected, there was a strong turnout from the telecom service providers who started the NFV initiative in late 2012, while many other companies were also represented such as network equipment manufacturers, system integrators, software companies and processor suppliers.

The overall impression was that key service providers are strongly committed to NFV, with the potential of interesting new business opportunities for a range of other players in an ecosystem that will have to include providers of hardware, software and services.

In terms of the business benefits of NFV, most of the discussion to date appears to be around the CAPEX and OPEX savings that can be achieved through the virtualization of functions that have traditionally been implemented as stand-alone, dedicated, fixed-function equipment.

One use case that’s frequently discussed is a “virtual CPE”, with an architecture that relies on a very simple CPE (modem, switch and antenna) with all services relocated to virtual network appliances on a central server. Another example is the cloud RAN (Radio Access Network) concept, in which the eNodeB at the antenna is replaced by a low-cost Remote Radio Head, with digitized RF data backhauled over fiber to a centralized Baseband Unit pool in which the baseband processing is performed. Finally, Evolved Packet Core (EPC) functions for mobile networks can be run in Virtual Machines (VMs) on generic server platforms.

Each of these scenarios promises to reduce both CAPEX and OPEX for service providers, though there’s still healthy debate as to whether a simple substitution of virtualized functions for dedicated equipment really reduces costs enough, once new operational expenses are factored in.

I was surprised not to hear more at the ESTI meeting about how NFV can enable new services, especially cloud-based applications, which would deliver real value to subscribers. These would seem to be the key to increasing Average Revenue per User (ARPU) and driving top-line growth in the P&L, rather than just reducing expenses through virtualization. On reflection, though, carriers probably view new services and applications as their proprietary secret sauce and not the kind of topic that they want to talk about in a room full of their competitors.

At this point in the evolution of NFV, the working groups and “expert groups” are mainly focused on topics relating to network management and orchestration topics. This makes perfect sense given the extreme complexity of legacy telecom networks and challenges of migrating these control-related systems to a completely new software-based architecture. And, of course, the concept of “Carrier Grade reliability” reflects expectations that we all have, as subscribers with the freedom to switch providers as soon as our contracts are up.

Before too long, though, we’ll need to talk about a fundamental data plane issue that must be addressed in order for any NFV implementation to be cost-effective.

Packet processing is a key function that dominates the processing workload for many telecom network subsystems. These include the CRAN and CPE solutions mentioned above, as well as several key EPC functions such as the Mobility Management Entity (MME), Packet Gateway, Security Gateway and Serving Gateway. In the case of traditional non-virtualized equipment, standard Operating System networking stacks provide poor performance for packet processing because of overheads and latencies within the kernel, so many equipment suppliers have adopted solutions such as 6WIND’s software, which solves that problem through a fast path architecture.

Virtualized implementations present additional challenges for packet processing. Virtualizing hundreds of CRAN, CPE or complex EPC functions on a single server requires high performance network bandwidth to be distributed to the virtualized functions. Standard virtualized architectures that work well for computing applications are unable to deliver the required performance for these demanding network applications. Bottlenecks in the hypervisor, the virtual switch and the VMs themselves can degrade overall networking performance by an order of magnitude for virtualized functions when compared to physical implementations. This is a potential showstopper in terms of the cost-effectiveness of NFV architectures, for which cost-per-subscriber will be a critical metric.

Fortunately, this problem can be solved. Through the use of techniques such as IO Virtualization (IOV), virtual NIC (vNIC) drivers and direct VM-to-VM (VM2VM) communication, as well as a fast path data plane implementation, it’s possible to bypass the bottlenecks. Advanced software solutions can deliver the high networking performance that’s required in order to achieve the cost reductions that are the basis of the NFV concept.

As an example of this, at Open Networking Summit during the week before the recent NFV meeting, 6WIND demonstrated a solution that delivers a 10x improvement in the performance of the standard Open Virtual Switch (OVS), which performs the critical switching function required to deliver high-bandwidth network traffic to VMs.

So while most of the NFV attention and conversations are currently centered on management and orchestration issues, let’s not forget that sooner or later we’ll need to make sure the data plane performance challenges are addressed too. This will be a critical factor in ensuring that the NFV architectures do indeed deliver the CAPEX and OPEX reductions that are promised. And then we’ll have low-cost, high-performance platforms for the delivery of innovative new services for subscribers, which can be the basic of increased service provider revenues.