Killing the ISP Appliance: An eBPF/XDP Approach to Distributed BNG

Summary

I used to work for an ISP startup that was building next-generation infrastructure. The company didn’t make it, but the problems we were trying to solve stuck with me. So I spent a few weeks building what we never got to: an open-source, eBPF-accelerated BNG that runs directly on OLT hardware.This post explains the architecture and why I think it’s the future of ISP edge infrastructure.The Problem: Centralised BNG is a BottleneckTraditional ISP architecture looks like this:Customer → ONT → OLT → [BNG Appliance] → Internet ↑ Single point of failure Expensive proprietary hardware All subscriber traffic flows through here Every subscriber’s traffic - DHCP, authentication, NAT, QoS - flows through a central BNG appliance. These boxes cost six figures, require vendor support contracts, and create a single point of failure. When they go down, everyone goes down.The industry’s answer has been to buy bigger boxes with more redundancy. But what if we flipped the model entirely?The Idea: Distribute the BNG to the EdgeWhat if, instead of funneling all traffic through a central appliance, we ran BNG functions directly on the OLT hardware at each edge site?Customer → ONT → OLT(+BNG) → Internet ↑ Subscriber traffic stays LOCAL No central bottleneck Each site operates independently This isn’t a new idea - it’s essentially what hyperscalers do with their edge infrastructure. But ISPs have been slow to adopt it because:Traditional BNG software assumes a central deploymentState management (IP allocations, sessions) is hard to distributePerformance requirements seemed to need specialised hardwareThe key insight is that modern Linux with eBPF/XDP can handle ISP-scale packet processing on commodity hardware.Why eBPF/XDP, Not VPP?When I started this project, I evaluated two approaches:VPP (Vector Packet Processing) - The industry darling for high-performance networking. Used in production by big telcos. Handles 100+ Gbps easily.eBPF/XDP - Linux kernel’s programmable packet processing. Lo...