Overview
This proposal outlines a design for a protocol enabling an I2P client, service or external balancer process to manage multiple routers transparently hosting a single [Destination].
The proposal currently does not specify a concrete implementation. It could be implemented as an extension to [I2CP], or as a new protocol.
Motivation
Multihoming is where multiple routers are used to host the same Destination. The current way to multihome with I2P is to run the same Destination on each router independently; the router that gets used by clients at any particular time is the last one to publish a [LeaseSet].
This is a hack and presumably won't work for large websites at scale. Say we had 100 multihoming routers each with 16 tunnels. That's 1600 LeaseSet publishes every 10 minutes, or almost 3 per second. The floodfills would get overwhelmed and throttles would kick in. And that's before we even mention the lookup traffic.
[Prop123] solves this problem with a meta-LeaseSet, which lists the 100 real LeaseSet hashes. A lookup becomes a two-stage process: first looking up the meta-LeaseSet, and then one of the named LeaseSets. This is a good solution to the lookup traffic issue, but on its own it creates a significant privacy leak: It is possible to determine which multihoming routers are online by monitoring the published meta-LeaseSet, because each real LeaseSet has corresponds to a single router.
We need a way for an I2P client or service to spread a single Destination across multiple routers, in a way that is indistinguishable to using a single router (from the perspective of the LeaseSet itself).
Design
Definitions
- User
- The person or organisation wanting to multihome their Destination(s). A single Destination is considered here without loss of generality (WLOG).
- Client
The application or service running behind the Destination. It may be a client-side, server-side, or peer-to-peer application; we refer to it as a client in the sense that it connects to the I2P routers.
The client consists of three parts, which may all be in the same process or may be split across processes or machines (in a multi-client setup):
- Balancer
- The part of the client that manages peer selection and tunnel building. There is a single balancer at any one time, and it communicates with all I2P routers. There may be failover balancers.
- Frontend
- The part of the client that can be operated in parallel. Each frontend communicates with a single I2P router.
- Backend
- The part of the client that is shared between all frontends. It has no direct communication with any I2P router.
- Router
- An I2P router run by the user that sits at the boundary between the I2P network and the user's network (akin to an edge device in corporate networks). It builds tunnels under the command of a balancer, and routes packets for a client or frontend.
High-level overview
Imagine the following desired configuration:
- A client application with one Destination.
- Four routers, each managing three inbound tunnels.
- All twelve tunnels should be published in a single LeaseSet.
Single-client
-{ [Tunnel 1]===\
|-{ [Tunnel 2]====[Router 1]-----
|-{ [Tunnel 3]===/ \
| \
|-{ [Tunnel 4]===\ \
[Destination] |-{ [Tunnel 5]====[Router 2]----- \
\ |-{ [Tunnel 6]===/ \ \
[LeaseSet]--| [Client]
|-{ [Tunnel 7]===\ / /
|-{ [Tunnel 8]====[Router 3]----- /
|-{ [Tunnel 9]===/ /
| /
|-{ [Tunnel 10]==\ /
|-{ [Tunnel 11]===[Router 4]-----
-{ [Tunnel 12]==/
Multi-client
-{ [Tunnel 1]===\
|-{ [Tunnel 2]====[Router 1]---------[Frontend 1]
|-{ [Tunnel 3]===/ \ \
| \ \
|-{ [Tunnel 4]===\ \ \
[Destination] |-{ [Tunnel 5]====[Router 2]---\-----[Frontend 2] \
\ |-{ [Tunnel 6]===/ \ \ \ \
[LeaseSet]--| [Balancer] [Backend]
|-{ [Tunnel 7]===\ / / / /
|-{ [Tunnel 8]====[Router 3]---/-----[Frontend 3] /
|-{ [Tunnel 9]===/ / /
| / /
|-{ [Tunnel 10]==\ / /
|-{ [Tunnel 11]===[Router 4]---------[Frontend 4]
-{ [Tunnel 12]==/
General client process
- Load or generate a Destination.
- Open up a session with each router, tied to the Destination.
- Periodically (around every ten minutes, but more or less based on tunnel
liveness):
- Obtain the fast tier from each router.
- Use the superset of peers to build tunnels to/from each router.
- By default, tunnels to/from a particular router will use peers from that router's fast tier, but this is not enforced by the protocol.
- Collect the set of active inbound tunnels from all active routers, and create a LeaseSet.
- Publish the LeaseSet through one or more of the routers.
Differences to I2CP
To create and manage this configuration, the client needs the following new functionality beyond what is currently provided by [I2CP]:
- Tell a router to build tunnels, without creating a LeaseSet for them.
- Get a list of the current tunnels in the inbound pool.
Additionally, the following functionality would enable significant flexibility in how the client manages its tunnels:
- Get the contents of a router's fast tier.
- Tell a router to build an inbound or outbound tunnel using a given list of peers.
Protocol outline
Client Router
---------------------> Create Session
Session Status <---------------------
---------------------> Get Fast Tier
Peer List <---------------------
---------------------> Create Tunnel
Tunnel Status <---------------------
---------------------> Get Tunnel Pool
Tunnel List <---------------------
---------------------> Publish LeaseSet
---------------------> Send Packet
Send Status <---------------------
Packet Received <---------------------
Messages
- Create Session
- Create a session for the given Destination.
- Session Status
- Confirmation that the session has been set up, and the client can now start building tunnels.
- Get Fast Tier
- Request a list of the peers that the router currently would consider building tunnels through.
- Peer List
- A list of peers known to the router.
- Create Tunnel
- Request that the router build a new tunnel through the specified peers.
- Tunnel Status
- The result of a particular tunnel build, once it is available.
- Get Tunnel Pool
- Request a list of the current tunnels in the inbound or outbound pool for the Destination.
- Tunnel List
- A list of tunnels for the requested pool.
- Publish LeaseSet
- Request that the router publish the provided LeaseSet through one of the outbound tunnels for the Destination. No reply status is needed; the router should continue re-trying until it is satisfied that the LeaseSet has been published.
- Send Packet
- An outgoing packet from the client. Optionally specifies an outbound tunnel through which the packet must (should?) be sent.
- Send Status
- Informs the client of the success or failure of sending a packet.
- Packet Received
- An incoming packet for the client. Optionally specifies the inbound tunnel through which the packet was received(?)
Security implications
From the perspective of the routers, this design is functionally equivalent to the status quo. The router still builds all tunnels, maintains its own peer profiles, and enforces separation between router and client operations. In the default configuration is completely identical, because tunnels for that router are built from its own fast tier.
From the perspective of the netDB, a single LeaseSet created via this protocol is identical to the status quo, because it leverages pre-existing functionality. However, for larger LeaseSets approaching 16 Leases, it may be possible for an observer to determine that the LeaseSet is multihomed:
The current maximum size of the fast tier is 75 peers. The Inbound Gateway (IBGW, the node published in a Lease) is selected from a fraction of the tier (partitioned randomly per-tunnel pool by hash, not count):
- 1 hop
The whole fast tier
- 2 hops
Half of the fast tier (the default until mid-2014)
- 3+ hops
A quarter of the fast tier (3 being the current default)
That means on average the IBGWs will be from a set of 20-30 peers.
In a single-homed setup, a full 16-tunnel LeaseSet would have 16 IBGWs randomly selected from a set of up to (say) 20 peers.
In a 4-router multihomed setup using the default configuration, a full 16-tunnel LeaseSet would have 16 IBGWs randomly-selected from a set of at most 80 peers, though there are likely to be a fraction of common peers between routers.
Thus with the default configuration, it may be possible through statistical analysis to figure out that a LeaseSet is being generated by this protocol. It might also be possible to figure out how many routers there are, although the effect of churn on the fast tiers would reduce the effectiveness of this analysis.
As the client has full control over which peers it selects, this information leakage could be reduced or eliminated by selecting IBGWs from a reduced set of peers.
Compatibility
This design is completely backwards-compatible with the network, because there are no changes to the [LeaseSet] format. All routers would need to be aware of the new protocol, but this is not a concern as they would all be controlled by the same entity.
Performance and scalability notes
The upper limit of 16 [Leases] per LeaseSet is unaltered by this proposal. For Destinations that require more tunnels than this, there are two possible network modifications:
- Increase the upper limit on the size of LeaseSets. This would be the simplest to implement (though it would still require pervasive network support before it could be widely used), but could result in slower lookups due to the larger packet sizes. The maximum feasible LeaseSet size is defined by the MTU of the underlying transports, and is therefore around 16kB.
- Implement [Prop123] for tiered LeaseSets. In combination with this proposal, the Destinations for the sub-LeaseSets could be spread across multiple routers, effectively acting like multiple IP addresses for a clearnet service.
Acknowledgements
Thanks to psi for the discussion that led to this proposal.
References
| [Destination] | http://i2p.net./spec/common-structures#struct-destination |
| [I2CP] | (1, 2) http://i2p.net./en/docs/protocol/i2cp |
| [Leases] | http://i2p.net./spec/common-structures#struct-lease |
| [LeaseSet] | (1, 2) http://i2p.net./spec/common-structures#struct-leaseset |
| [Prop123] | (1, 2) http://i2p.net./spec/proposals/123-new-netdb-entries |