VIRTUAL BROADBAND NETWORK GATEWAY

A vBNG is a BNG that runs on a server you bought.

That is the whole definition. Everything interesting is in the second question: whether the software running on that server does the same job as the chassis it is replacing, at the rate your network actually needs.

The short answer: a vBNG (virtual Broadband Network Gateway) is broadband network gateway software running on general-purpose x86 hardware instead of a purpose-built chassis. It terminates the same subscriber sessions — PPPoE and IPoE — does the same AAA, QoS and IPv6, and connects to the same RADIUS. What changes is the supply chain: you buy servers, and capacity is a software licence rather than a line card.

The "v" is historical. It came from network-function virtualisation, but most vBNGs in production today run on bare metal, because a hypervisor between the packet and the network card costs exactly the thing the box is sold on.

Where the “v” came from, and why it is misleading

The term arrived with network-function virtualisation: the idea that functions which had always shipped as dedicated boxes — firewalls, routers, broadband gateways — could be packaged as software and consolidated onto commodity servers. For a lot of functions that worked well.

For a data plane it worked less well, and the reason is mechanical. Every layer between the network card and the forwarding code costs per-packet time, and a broadband gateway is measured in packets per second. A hypervisor, a virtual switch and a virtual NIC are three such layers. Operators who cared about the number ended up running the software directly on the hardware.

So the honest reading of “vBNG” today is software BNG. If a vendor is still selling you a VM image and quoting throughput, ask what it was measured on, and at what frame size.

What has to be true before it is a real BNG

Plenty of software can terminate a PPPoE session. That is not the same as being able to replace the box in your rack. Six things separate the two:

Both session types, together

PPPoE and IPoE on the same box at the same time, because real access networks are mid-migration for years at a stretch.

AAA that plugs into what you run

Standards-based RADIUS with accounting and CoA — not a proprietary subscriber database you now have to keep in step with your billing.

Scheduling, not a rate limit

Hierarchical QoS: subscriber inside VLAN inside port. A single token bucket per subscriber is not the same product.

Sessions as a scale number

Throughput is the easy number. Ask how many concurrent sessions, how fast they re-establish after an outage, and what memory that costs.

Packet rate with everything on

Quoted with the full feature set enabled and the frame size stated, or the figure cannot be checked and is not worth having.

Someone who owns the driver

At line rate the fault is often below the application. The useful question is who fixes the network-card driver when it is the driver.

The performance comes from one decision

Everything above is table stakes. The number on the datasheet comes almost entirely from where the forwarding code runs, and there are three answers:

The ordinary kernel path. Simple, and fine for small deployments. Every packet gets a socket buffer allocated and walks the full network stack, which sets the ceiling early.
Kernel bypass. A poll-mode driver takes the network card away from the kernel and forwards in userspace. Fast, at the cost of dedicated cores spinning whether or not traffic is arriving, reserved memory, and a second routing table to keep in step with the first.
In-kernel, in the driver. An eBPF program runs in the network driver before the socket buffer is allocated. The port stays a normal interface, nothing is reserved in advance, and processor cost tracks offered load.

BNGSOFT took the third. The XDP data plane page sets out what that buys and what it costs, with the measured numbers and the arithmetic behind them.

“Is there an open-source vBNG?”

People search for this often enough that it deserves a straight answer: there are excellent open-source components, and there is no open-source product that replaces a carrier BNG on its own.

The building blocks are real. accel-ppp is a mature PPPoE, PPTP, L2TP and IPoE access concentrator. FRRouting handles BGP and the rest of the routing. VPP gives you a high-performance userspace data plane. Linux itself brings the traffic control subsystem, netfilter and, more recently, eBPF and XDP.

What none of them gives you is the integration: one subscriber identity flowing from RADIUS through the scheduler, the NAT table and the logging; one upgrade that does not drop sessions; one place to look when a customer calls. That integration, and someone whose job it is to fix the layer below when it breaks, is what a vendor is actually selling. Anyone telling you the components are the product has not run one.

What actually changes when you move off a chassis

	Purpose-built chassis	Software BNG on x86
Adding capacity	A line card, and a slot to put it in	Another server, or a bigger licence on the one you have
Lead time	Quoted in months	Whatever your server supplier quotes
Failure domain	One large box, redundant internally	Several smaller boxes, redundant by count
Upgrades	Maintenance window, vendor release train	In-service, on your own schedule
Security fixes	When the next image ships	As software, in days
Sparing	Vendor-specific, held by you or them	Any identical server
The hard part	Cost per subscriber	Proving the packet rate on your own traffic

Neither column is automatically right. A chassis is still the correct answer in plenty of places — usually where the port density, not the subscriber count, is what you are buying.

Go deeper

Prove the packet rate on your own traffic.

Tell us subscriber count and busy-hour throughput and we will size it against your numbers. A proof of concept runs on your own x86, against your own traffic.