The I2P Network Protocol (I2NP), which is sandwiched between I2CP and the various I2P transport protocols, manages the routing and mixing of messages between routers, as well as the selection of what transports to use when communicating with a peer for which there are multiple common transports supported.
I2NP (I2P Network Protocol) messages can be used for one-hop, router-to-router, point-to-point messages. By encrypting and wrapping messages in other messages, they can be sent in a secure way through multiple hops to the ultimate destination. Priority is only used locally at the origin, i.e. when queuing for outbound delivery.
The priorities listed below may not be current and are subject to change. See the OutNetMessage Javadocs for the current priority settings. Priority queueing implementation may vary.
The following table specifies the traditional 16-byte header used in NTCP. The SSU and NTCP2 transports use modified headers.
|Nutzdaten||0 - 61.2KB|
While the maximum payload size is nominally 64KB, the size is further constrained by the method of fragmenting I2NP messages into multiple 1KB tunnel messages as described on the tunnel implementation page. The maximum number of fragments is 64, and the message may not be perfectly aligned, So the message must nominally fit in 63 fragments.
The maximum size of an initial fragment is 956 bytes (assuming TUNNEL delivery mode); the maximum size of a follow-on fragment is 996 bytes. Therefore the maximum size is approximately 956 + (62 * 996) = 62708 bytes, or 61.2 KB.
In addition, the transports may have additional restrictions. The NTCP limit is 16KB - 6 = 16378 bytes. The SSU limit is approximately 32 KB. The NTCP2 limit is approximately 64KB - 20 = 65516 bytes, which is higher than what a tunnel can support.
Note that these are not the limits for datagrams that the client sees, as the router may bundle a reply leaseset and/or session tags together with the client message in a garlic message. The leaseset and tags together may add about 5.5KB. Therefore the current datagram limit is about 10KB. This limit will be increased in a future release.
Higher-numbered priority is higher priority. The majority of traffic is TunnelDataMessages (priority 400), so anything above 400 is essentially high priority, and anything below is low priority. Note also that many of the messages are generally routed through exploratory tunnels, not client tunnels, and therefore may not be in the same queue unless the first hops happen to be on the same peer.
Also, not all message types are sent unencrypted. For example, when testing a tunnel, the router wraps a DeliveryStatusMessage, which is wrapped in a GarlicMessage, which is wrapped in a DataMessage.
|DatabaseSearchReplyMessage||3||Typ. 161||300||Size is 65 + 32*(number of hashes) where typically, the hashes for three floodfill routers are returned.|
|DatabaseStoreMessage||1||Varies||460||Priority may vary. Size is 898 bytes for a typical 2-lease leaseSet. RouterInfo structures are compressed, and size varies; however there is a continuing effort to reduce the amount of data published in a RouterInfo as we approach release 1.0.|
|DataMessage||20||4 - 62080||425||Priority may vary on a per-destination basis|
|DeliveryStatusMessage||10||12||Used for message replies, and for testing tunnels - generally wrapped in a GarlicMessage|
|GarlicMessage||11||Generally wrapped in a DataMessage - but when unwrapped, given a priority of 100 by the forwarding router|
|TunnelDataMessage||18||1028||400||The most common message. Priority for tunnel participants, outbound endpoints, and inbound gateways was reduced to 200 as of release 0.6.1.33. Outbound gateway messages (i.e. those originated locally) remains at 400.|
|VariableTunnelBuildMessage||23||1057 - 4225||500||Shorter TunnelBuildMessage as of 0.7.12|
|VariableTunnelBuildReplyMessage||24||1057 - 4225||300||Shorter TunnelBuildReplyMessage as of 0.7.12|
|Others listed in 2003 Spec||0,4-9,12||Obsolete, Unused|
Full Protocol Specification
It isn't clear whether the current priority scheme is generally effective, and whether the priorities for various messages should be adjusted further. This is a topic for further research, analysis and testing.