Spanner-RSS

What is Spanner-RSS?

Spanner-RSS is a variant of Spanner, Google's globally distributed database, that relaxes its consistency from strict serializability to regular sequential consistency. In doing so, Spanner-RSS is able to offer lower tail latency than Spanner for read-only transactions. This code was used for the SOSP 2021 paper, "Regular Sequential Serializability and Regular Sequential Consistency." It is based off of the code originally used in the evaluation of TAPIR from SOSP 2015.

This repository includes an implementation of Spanner's protocol (as described in Google's 2013 TOCS paper), an implementation of our Spanner-RSS variant, and scripts to run the experiments presented in our paper.

Compiling & Running

Tool Versions

We've built and run the spanner-rss with the following compiler tools:

• cmake v3.10.2
• gcc/g++ v7.5.0
• python v3.6.9
• gnuplot v5.2

Running experiments

Experiments for the paper were run on CloudLab using the spanner-rss profile. After starting an experiment,

1. Clone the experiment repository to one of the CloudLab machines. (We often use client-0-0.)
   $ git clone --recursive https://github.com/princeton-sns/spanner-rss.git

2. Build the C++ benchmark and server.
   $ cd spanner-rss/src/ && mkdir build && cd build
$ cmake .. && make

3. Install experiment script dependencies.
   $ sudo apt update && sudo apt install -y python3-numpy gnuplot

4. Update experiment config. You will likely need to update the following fields:

   • project_name
   • experiment_name
   • base_local_exp_directory
   • base_remote_bin_directory_nfs
   • src_directory
   • src_commit_hash

5. After updating the config file, you can run the experiment using a python3 script. For example,
   $ python3 ./experiments/run_multiple_experiments.py experiments/configs/retwis-wan-tput-lat-zipf-0.7.json

Authors

Jeffrey Helt, Amit Levy, Wyatt Lloyd -- Princeton University

Matthew Burke -- Cornell University