Prefix translation and advanced multihoming
Explain NPTv6 (experimental) and its trade-offs (RFC 6296). Compare NPTv6 with provider-assigned (PA) multihoming without translation (RFC 8678). Recommend a design for a small multihomed site.
Take the quiz · pass mark 70% · 9 questions
Explanation
Multihoming to two ISPs with IPv6 raises a classic question: how do internal hosts use two provider prefixes without the pain of running BGP? There are two schools — translate (NPTv6) or don't translate (PA multihoming).
NPTv6 (RFC 6296, experimental)
NPTv6 (IPv6-to-IPv6 Network Prefix Translation) is a stateless, 1:1 mapping between an internal prefix and an external prefix. It rewrites only the prefix part of the address and leaves the interface identifier untouched, and it is checksum-neutral (it adjusts bits so the transport checksum is unchanged, so there is no need to touch Layer 4). It is explicitly marked experimental and is not classic stateful NAT: there is no address/port overloading and no connection table. For a /48-to-/48 mapping it translates the top 48 bits and preserves the lower 80 (including the 64-bit IID), so a host keeps a stable internal address while appearing under whichever provider prefix it egresses.
The appeal: internal hosts use one stable (often ULA or PA) prefix, and NPTv6 presents them under each ISP's prefix at the edge, so a provider change or second provider does not renumber the inside. The cost: it breaks the IPv6 end-to-end model, complicates anything that embeds addresses (referrals, logging, some security), and is experimental — the IETF did not bless it as the recommended multihoming answer.
PA multihoming without translation (RFC 8678)
RFC 8678 describes enterprise multihoming using provider-assigned (PA) addresses without translation: hosts receive both providers' prefixes (multiple addresses per interface, Level 2), and the network uses source-address-dependent routing plus correct RFC 6724 address selection and RAs (with per-prefix lifetimes) so a host picks the right source for the exit it will use. No translation, so the end-to-end model is intact; the complexity moves into routing and address selection rather than a translator.
Choosing for a small site
- Small site, wants simplicity and no renumbering, can tolerate experimental translation: NPTv6 is tempting and operationally simple, but weigh the experimental status and end-to-end breakage.
- Site that values the end-to-end model and clean operations: prefer PA multihoming (RFC 8678) with source-dependent routing, accepting multiple addresses per host.
- Larger site: PI space with BGP (Level 4) is the real answer; translation is a workaround for those who cannot run BGP.
The honest recommendation mirrors the IETF's: prefer no translation where you can; use NPTv6 only with eyes open.
ictNetworks scenario
For a small dual-homed business customer that cannot run BGP, ictNetworks offers PA multihoming (RFC 8678) as the default — both prefixes advertised, source-dependent routing and RFC 6724 selection — and only falls back to NPTv6 (RFC 6296) when the customer insists on a single stable internal prefix and accepts the experimental trade-offs. It documents that NPTv6 breaks the end-to-end model and keeps it off the main enterprise design, which uses PI + BGP (L4.06).
Drawing diagram…
Worked examples
Example 1: What NPTv6 rewrites
For a /48-to-/48 NPTv6 mapping, the translator rewrites the top 48 bits (the prefix) and leaves the lower bits unchanged, so only the network part changes as the packet crosses the edge.
Example 2: What it preserves
NPTv6 leaves the 64-bit interface identifier untouched, so a host keeps a stable, recognisable lower half of its address regardless of which provider prefix it appears under.
Example 3: Prefixes for a dual-ISP site
A dual-ISP NPTv6 site maps its single internal /48 to 2 external provider prefixes (one per ISP), choosing which to present based on the egress.
Lab
Goal: Implement NPTv6 for a dual-ISP site and compare it with the L4.06 design.
You need: an nftables-based edge with two simulated ISP uplinks, an internal LAN on a stable prefix, and test hosts. (nftables NPTv6 support and syntax vary by version — check the docs.)
- Number the internal LAN from a stable prefix (ULA or PA) and bring up two ISP uplinks with documentation prefixes.
- Configure NPTv6 (RFC 6296) mapping the internal /48 to each ISP's /48, checksum-neutral; verify outbound traffic appears under the correct provider prefix.
- Confirm the interface identifier is preserved and the mapping is stateless (no connection table).
- Demonstrate a break case: an application that embeds its own address (or a scenario needing end-to-end address integrity) and show how NPTv6 complicates it.
- Contrast with the L4.06 PA/PI multihoming design (no translation) and write up when you would choose each (RFC 8678).
What to submit: the nftables NPTv6 configuration, evidence of per-provider prefix presentation with preserved IID, the break-case demonstration, and the NPTv6-vs-no-translation comparison.
Summary
- NPTv6 (RFC 6296, experimental) is stateless, checksum-neutral 1:1 prefix translation that rewrites the prefix and keeps the IID.
- It avoids renumbering but breaks the end-to-end model and complicates address-embedding apps.
- PA multihoming without translation (RFC 8678) keeps the end-to-end model using multiple prefixes, source-dependent routing and RFC 6724 selection.
- Prefer no translation where you can; for larger sites use PI + BGP (L4.06). Use NPTv6 only with eyes open.
- Next lesson (L8.05): anycast services and large-scale design.
Quiz imported · 9 questions · pass mark 70%. Lab questions are not auto-marked.