|Nov 1, 2023
|May 17, 2023
|Nov 11, 2022
|Jul 27, 2022
|Nov 2, 2021
#1 in #kad
1,048 downloads per month
Used in dusk-consensus
Implementation of the Kadcast Network layer according to the latest version of the paper to date (2021)
Kadcast is an UDP-based peer-to-peer protocol in which peers form a structured overlay.
💡 The specifications suggest that the protocol can be transport-agnostic, however this implementation uses the canonical choice of UDP.
All peers have an unique ID generated upon joining the network, which is used as an identifier unique to them.
💡 This implementation uses 128 bits identifiers generated from hashing the
<ip_octects>concatenation truncated to the agreed 16 bytes. The hashing algorithm used is
Following the Kademlia's specifications for a structured network, the ID determines the peer position in a binary routing tree.
Each peer mantains the routing state in a set of
k-buckets containing the location of the location of each peer. The location is fully determined by the tuple
(ip_addr, port, ID).
Each bucket comprises of a list of the
K last recently seen peers, located at a given distance in relation to the peer. The distance is calculated as the
non-euclidean XOR metric which represents:
d(x,y) = (x ^ y).
This implies that
i buckets contain
K Peers which distance
d follow the formula:
2^i <= d < 2^(i +1).
💡 The factor
Kis a system-wide parameter which determines the routing state and the lookup complexity.
In case a bucket's peer list is full (i.e.
kentries are already stored), new entries to that list may be inserted by dropping
LRU(Least Recently Used) peers.
Before dropping an entry from the list, the peer will check whether the peer to be dropped is still reachable, by sending a
PINGmessage and waiting for a
PONGresponse. In case the
PONGwon't be collected within pre-determined timeout, the check fails and the peer is dropped from the list.
💡 By following a
LRUpolicy, the protocol inherently is biased toward the older and more stable peers on the network.
Upon joining the Kadcast network, a peer retrieves from the so-called
bootstrap nodes the locations of other neighbour peers, so to fill its buckets.
💡 From the specification: When a peer first joins the network, it has to know the address of at least one bootstrapping node. It therefore sends
PINGmessages to known peers to check whether they are actually online. Additionally,
PINGtransmits the sending peer’s routing information to the recipient, thereby distributing its existence in the network.
The bootstrap node will reply with a
PONG message which also contains its node info.
Starts just after the bootstrapping process.
The main goal of this phase is to fill up the
k different peers.
- The process starts by sending a
FIND_NODEmessage to the bootstraping nodes that have been added to the
k-buckets with the
- Then the lookup process starts:
- The node looks up the 𝛼 closest peers regarding the XOR-metric in its own buckets.
- It queries this 𝛼 peers for the ID by sending
FIND_NODEmessages and wait for the
- The queried peers respond with a
NODESreply containing the set of
kpeers they believe are the closest to
- Based on the info acquired, the node builds a new set of closest peers and repeats steps 1 to 3 until it no longer gets peers closer to those it got from previous iterations.
⚠️ Note that like the bucket size,
kare global constants that will determine the redundancy and overhead of the network. The typical values are:
k = [20, 100]&
𝛼 = 3.
Periodic Network Manteinance
Each peer periodically refreshes every bucket with no activity. For each such bucket, it picks a random
ID with appropriate distance, performs a look up, and populates its buckets with fresh routing information.
Improving Broadcast Reliability And Performance
Instead of delegating the broadcast to one peer per bucket, we select
β delegates. This is done to increase the probability of at least one honest peer to receive the message broadcasted. This is a measure intended to help with the inherently lossy nature of the UDP protocol without forcing the use of slow CRC checking procedures.
Since each broadcast process is repeated on every hop (decreasing the height), it is imperative that the peers ignore duplicate
CHUNK messages. Moreover, transmission failures have to be considered. Because of these factors, this implementation includes the use of forward error connection schemes based on RaptorQ specifications and related rust library.
FEC overhead factor can be adjusted through the paramenter
f = (n-s)/s.
As shown in the benchmarks, we can get full network coverage even assuming a 12% packet-loss ratio with a
β = 3 &
f = 0,15 (Please refer to the benchmarks included with the library documentation).
The specification provides solutions for DOS, Sybil and Eclipse attacks, as well as obstruction of block delivery. All advices have been taken into consideration during the development.
For more information related to the internal architecture please check the architecture diagram.