This repository contains the prototype implementation of SwiftPaxos, a new state-machine replication protocol for geo-distributed systems. SwiftPaxos is a faster Paxos without compromises. In the best case, it executes a state-machine command in two message delays (one round-trip), and three otherwise. SwiftPaxos was presented at the 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24).
git clone https://github.com/imdea-software/swiftpaxos.git
cd swiftpaxos
go install github.com/imdea-software/swiftpaxos
Protocol | Comments |
---|---|
SwiftPaxos | See our NSDI'24 paper for the full details. |
Paxos | The classic Paxos protocol. |
N2Paxos | All-to-all variant of Paxos. |
CURP | CURP implemented over N2Paxos. |
Fast Paxos | Fast Paxos with uncoordinated collision recovery. |
EPaxos | A corrected version of EPaxos. |
This software is based on the Egalitarian Paxos code base, as well as the corrections made here.
There are three types of participants: master, servers and clients. The servers and clients implement the protocol logic. The master maintains the configuration of the system.
To setup a run, the participants read deployment configuration file. See aws.conf for an example of configuration file for AWS EC2.
Master:
swiftpaxos -run master -config conf.conf
Server:
swiftpaxos -run server -config conf.conf -alias server_name
Client:
swiftpaxos -run client -config conf.conf -alias client_name
-alias alias
An alias of this participant
-config file
Deployment config file (required)
-latency file
Latency config file
-log file
Path to the log file
-protocol protocol
Protocol to run. Overwrites protocol field of the config file
-quorum file
Quorum config file
-run participant
Run a participant
See quorum.conf and latency.conf for an example of quorum and latency configuration files.
To have an idea on how different replication protocols would compare, we wrote a tool named Flint. Flint takes as input a set of AWS regions. It computes the expected latencies and estimates how the protocols perform in such a deployment.