This project presents a simple implementation of Paxos in Python, and several workloads that are used to measure its performance with different configurations. In particular, we will be measuring the performance of the algorithm when the ledgers must be written to a durable and reliable storage (such as the disk), and when it can remain in RAM.
For the architecture, check out /writeup/final_paper.pdf. The core code can be found in /paxos/.
This code has been tested on both Mac OS X El Capitan and the latest release of Ubuntu.
Clone this repository, and use Python on any of the workloads in /workloads/ or the tests in /paxos/test/ in order to test the implementation. Documentation on each workload/test is provided
inside its .py file.
This repository includes a compiled version of the Protocol Buffers used in gRPC. If you want to compile it, follow the instructions in
https://developers.google.com/protocol-buffers/
In order to run the gRPC version of the code, you will also need to install the gRPC library for Python:
pip install grpcio
This assumes you have pip installed; this is easy to get online.
Once you start up a Paxos workload (i.e. with python sample_workload.py) it will spawn a network of virtual machines talking to each other and passing greek decrees, but it will, by default, print nothing of what the Paxos agents are actually choosing. In order to make them actually print what they are doing you can change the DEBUG flags in proposer.py, learner.py, and acceptor.py. All of that debug output, however, is very verbose.
If you wish to simply see the ledger of one of the learners, you can run the client.py found in workloads/. It accepts arguments of the type
host:port:command
sending a given command to the VM running at (host, port). The possible commands are
print_ledger
print_differences
print_difference_mean
Where print_ledger prints all of the passed decrees learned by the given VM. The other two commands are used to print the differences in time between the time a proposal is sent and the time the decree is passed.
A common usage would be to spawn a workload with some VM running in port 6666 in a terminal window, and then open the client in another window and invoke
localhost:6666:print_ledger
Please note that, as of now, the Paxos client only works with systems using the RDTP message passing mechanism. Currently, the client cannot send requests to VMs using Google RPC.
Documentation for the code can be found in the /doc/ directory. It contains
HTML compiled by Sphinx. The repository already includes the compiled version
of the documentation. The main documentation HTML file is
/doc/index.html
and you can navigate through the nice Sphinx interface easily.