VPN architectures: from two tunnels to an interconnected estate
A tunnel joins two points; an estate needs a topology, and that topology is the design.

A virtual private network (VPN) tunnel joins two points; an estate of sites needs a topology of them, and that topology, not the encryption, is the architecture. Two sites need one tunnel and no decision; the third forces a choice: a tunnel to every other site, or a centre. It settles what the routing protocol sees, what carries the estate when a tunnel drops, and whether the sites’ addresses overlap. NIST SP 800-77 Revision 1 names four primary IPsec architectures: gateway-to-gateway, remote access, host-to-host and mesh (nist-sp800-77r1.txt lines 1110–1123).
Two tunnels are not yet an architecture
Two sites, one tunnel: no architecture to choose. It begins at the third, where tunnels that share no policy are separate problems. NIST calls the fully connected option “mesh encryption” (nist-sp800-77r1.txt lines 1315–1318); it is the costliest: four sites need six tunnels, ten need forty-five, each state to configure, key and keep alive.
Hub-and-spoke and full mesh
Hub-and-spoke replaces that with a centre: every site has one tunnel to the hub, spoke-to-spoke traffic transits it, and the count grows with the sites, not their square. A spoke-to-spoke flow takes two tunnels, and the hub becomes a single point of failure carrying every inter-site flow. NIST warns that “The architecture should take into consideration the effects of IPsec gateway outages” (nist-sp800-77r1.txt lines 3556–3560).
hub-and-spoke full mesh
A B A ────── B
\ / │ \ / │
\ / │ \ / │
[ H ] │ \/ │
/ \ │ /\ │
/ \ │ / \ │
C D C ────── D
The tunnel as a link for routing
To the routing protocol a tunnel is an interface with a metric. “Routing-based IPsec accepts all traffic at the IPsec policy layer, but both endpoints filter valid traffic by setting routes into a specific IPsec interface”, so “the routing table acts as the policy filter” (nist-sp800-77r1.txt lines 938–940). The route into the tunnel points at its far end, reached by the path underneath: one table carries two layers. A route is also weaker than a policy: “using routes can be fragile”, NIST warns, because “Another subsystem could change the routing to accidentally or maliciously bypass the IPsec device” (lines 3299–3301). Overhead is the second cost: stalls follow from “the extra overhead of the ESP header making each 1500-byte original packet larger than 1500 bytes after the ESP header is added” (lines 3342–3344), remedied with a smaller tunnel MTU or a clamped TCP segment size (lines 3356–3357).
Detecting the failure the tunnel cannot report
The tunnel does not tell the routing protocol its path is gone. Routing protocols’ own “relatively slow "Hello" mechanisms” leave failures invisible longer than a busy link affords (rfc5880.txt lines 141–145), and “BFD can provide failure detection on any kind of path between systems, including direct physical links, virtual circuits, tunnels” (lines 205–208). Redundancy — a second hub, or a second tunnel ranked by metric — is only as good as its independence: NIST requires that “if the network is designed to be redundant, the IPsec gateways and management servers should also be designed to be redundant” (nist-sp800-77r1.txt lines 3569–3571).
Handing the interconnection to a provider
A provider that builds the interconnection changes the vocabulary, not the question. CE-based: “the shared service provider network does not have any knowledge of the customer VPN” (rfc4026.txt lines 367–372). PE-based: “the PE device maintains VPN state, isolating users of one VPN from users of another” (lines 376–381). The blocks are the customer edge (CE), the provider edge (PE) that keeps VPN state, and the VPN-unaware provider core router (P) (rfc4026.txt lines 432, 516, 601–603). Services follow the layer: an L3VPN “interconnects sets of hosts and routers based on Layer 3 addresses” (lines 192–193); a layer 2 service emulates a wire or a LAN between customer edges (lines 206–207, 217–222).
The idea behind BGP/MPLS IP VPNs
BGP/MPLS IP VPNs are the reference design for provider-managed layer 3 interconnection; their claim is negative: the sites are not an overlay. Customer edge routers do not peer, so “there is no "overlay" visible to the VPN’s routing algorithm” (rfc4364.txt lines 132–133), and the provider core carries no per-VPN routes, forwarding on labels alone (lines 378–382). Separation is by instance — one forwarding table per site, “a per-site forwarding table” (rfc4026.txt lines 873–880) — and a route target decides which instances receive it (rfc4364.txt lines 850–851). Overlapping addresses are why: two sites in no common VPN “may have overlapping address spaces” (lines 327–332), and a route distinguisher — “an 8-byte value that, together with a 4 byte IPv4 address, identifies a VPN-IPv4 address family” (rfc4026.txt lines 793–798) — lets one table hold two routes to a prefix, one per VPN (rfc4364.txt lines 683–689).
The common mistake: an overlay is not a zone
Interconnecting sites is transport, not a boundary. Provider-managed layer 3 VPNs “do not by themselves encrypt the data for privacy, nor do they provide a way to determine whether data has been tampered with en route” (rfc4364.txt lines 410–413), and the isolation they give is between the provider’s customers only (lines 405–409); the sibling sheet on site-to-site and remote-access VPNs makes the same point for tunnels. A full mesh sharpens it: “A full mesh encryption bypasses all network-based firewalls and intrusion detection systems” (nist-sp800-77r1.txt lines 3544–3546).
Level and prerequisites
L4. It assumes the L3 sheet on site-to-site and remote-access VPNs (the tunnel, the IPsec elements, the trust model) and the L2 material on routing and the next hop.
Where to go next
References
- RFC 4026 — Provider Provisioned Virtual Private Network (VPN) Terminology (Andersson, Madsen, March 2005, Informational): the CE/PE/P building blocks, the CE-based and PE-based families, the L2 and L3 services, and the route distinguisher, route target and VRF vocabulary.
- RFC 4364 — BGP/MPLS IP Virtual Private Networks (VPNs) (Rosen, Rekhter, February 2006, Standards Track): the peer model and the absent overlay, the provider core with no per-VPN routes, route targets, the VPN-IPv4 address family, overlapping address spaces, and the security a provider-managed VPN does not provide.
- NIST SP 800-77 Revision 1 — Guide to IPsec VPNs (June 2020): the four primary IPsec architectures, mesh encryption, routing-based versus policy-based VPNs, the tunnel’s overhead and MTU, and gateway outages.
- RFC 5880 — Bidirectional Forwarding Detection (BFD) (Katz, Ward, June 2010, Standards Track): failure detection on tunnels and the limits of routing-protocol hellos.