mweidner037 / fugue

Fugue list CRDT implementations and benchmarks

Fugue code and benchmarks

Code and benchmarks for the paper "The Art of the Fugue: Minimizing Interleaving in Collaborative Text Editing" by Matthew Weidner and Martin Kleppmann, 2023. See the paper's Section 5 for a description.

Instructions

To reproduce the paper benchmarks:

1. Install Node.js (we used v18.15.0, but newer should also work).
2. In this folder:

# Install dependencies.
npm ci
# Build fugue implementations, run most benchmarks, and rewrite results_table.md.
# Takes ~2 hours.
npm start
# Run real text trace x100 benchmarks and rewrite results_table.md.
# Takes ~3 hours.
npm run B4x100

Statistics are written to results_table.md, while raw data is written to crdt-benchmarks/benchmarks/results.json. Those files initially contain the measurements used in the paper.

results.json is additive: each benchmark run adds its data to the existing file, only overwriting previous data for the same benchmarks. results_table.md is generated from the full results.json each time. To delete both files and start fresh, do npm run dataclean.

To speed things up, you can reduce the _TRIALS counts in crdt-benchmarks/js-lib/utils.js.

Implementations

• fugue/: Fugue
• fugue-simple/: Fugue Simple
• fugue-max-simple/: FugueMax Simple

Benchmarks info

The crdt-benchmarks folder is modified from https://github.com/dmonad/crdt-benchmarks/ by Kevin Jahns. We added our Fugue implementations to the suite (in crdt-benchmarks/benchmarks), removed measurements that are not relevant to the paper, and made minor tweaks (e.g. in B4, we measure memory usage of the "live" typing user instead of a future user who loads the document).

Info on microbenchmarks (measured but not described in the paper) is found in crdt-benchmarks/README.md.

Misc

yjs-interleave/ runs the Yjs backward interleaving (multi-replica) code from the paper's Appendix A.1.9. Run with npm start in that folder; the code is in yjs-interleave/index.js.

charts.xlsx is a spreadsheet used to calculate average throughputs and the over-time charts.