Last updated


This proposal adds an I2CP option for outbound tunnels that causes tunnels to be picked or built when a message is sent such that the OBEP matches one of the IBGWs from the [LeaseSet] for the target [Destination].


Most I2P routers employ a form of packet-dropping for congestion management. The reference implementation uses a WRED strategy that takes both message size and travel distance into account [TUNNEL-THROTTLING]. Due to this strategy, the primary source of packet loss is the OBEP.


When sending a message, the sender picks or builds a tunnel with an OBEP that is the same router as one of the recipient's IBGWs. By doing so, the message will go directly out of one tunnel and into the other, without needing to be sent across the wire in between.

Security implications

This mode would effectively mean that the recipient is selecting the sender's OBEP. In order to maintain current privacy, this mode would cause outbound tunnels to be built one hop longer than specified by the outbound.length I2CP option (with the final hop possibly being outside the sender's fast tier).


A new I2CP option is added to [I2CP-SPEC]:



Default value: case-specific

If true, the router will pick outbound tunnels for messages sent during this session such that the tunnel's OBEP is one of the IBGWs for the target Destination. If no such tunnel exists, the router will build one.


Backwards-compatibility is assured, as routers can always send messages to themselves.


Java I2P

Tunnel building and message sending are currently separate subsystems:

  • BuildExecutor only knows about the outbound tunnel pool's outbound.* options, and has no visibility regarding their use.
  • OutboundClientMessageOneShotJob can only select a tunnel from the existing pool; if a client message comes in and there are no outbound tunnels, the router drops the message.

Implementing this proposal would require designing a way for these two subsystems to interact.


A test implementation has been completed.


This proposal has various effects on latency, RTT and packet loss:

  • It is likely that in most cases, this mode would require building a new tunnel on first message rather than using an existing tunnel, adding latency.
  • For standard tunnels, the OBEP may need to find and connect to the IBGW, adding latency that increases the first RTT (as this occurs after the first packet has been sent). Using this mode, the OBEP would need to find and connect to the IBGW during tunnel building, adding the same latency but reducing the first RTT (as this occurs before the first packet has been sent).
  • The currently-standard [VariableTunnelBuild] size is 2641 bytes. Thus it is expected that this mode would result in lower packet loss for average message sizes larger than this.

More research is necessary to investigate these effects, in order to decide which standard tunnels would benefit from this mode being enabled by default.