Atomic broadcast with Paxos in Rust.
This program implements a distributed programming primitive called Atomic Broadcast using the Paxos [1] algorithm in the Rust programming language. The program uses IP multicast as communcation method between different processes, each one of the Paxos "actors". It can be run on a single machine.
Paxos is implemented with the optimisation where 2B messages are sent directly to the learner. A timeout is used to restart incomplete instances where message loss occurred.
- Install Rust.
- Clone repository and compile:
git clone https://github.com/daviderovell0/paxos-rs && cd paxos-rs
cargo build- Run tests via the scripts:
tests/run.sh tests <N> # vanilla-run
tests/run_loss.sh tests <N> # nThe script will generate N integers, feed them to the client processes
and send them to the Paxos proposers, starting the paxos instance.
Each Paxos process (proposer, acceptor, learner) is started separately
via the corresponding scripts in tests.
Debugging info and message being sent are written to stdout.
Values proposed and learned are written to prop<n> and learn<n> files
respectively.
All values proposed are learned correctly in the vanilla run for an arbitrary number of values proposed. To test with more than 1000 values increase the last sleep timeout in the testing scripts.
When message losses occur, messages that are seen by the proposer are learned most of the time. Note that messages that fail to reach the proposer (sent by clients and lost) are not considered by Paxos. Current limitation is that messages that are lost after phase 2A is sent (proposer-> acceptor) might not be recovered, as the restart thread monitor only the proposer.
No learner-catch up mechanism is implemented.
This program was developed as possible solution for the final assignment of the Distributed Algorithms 2022 course at USI held by professor Fernando Pedone, Eliã Rafael de Lima Batista and Nenad Milosevic.
- [1] Lamport, Leslie. “Paxos Made Simple.” (2001). link