VLDB 2024: Boot

2024-09-09: This repository is a proof-of-concept extracted from the Database Gym project. If you're looking to use Boot, especially for PostgreSQL, check that out instead.

---

Source code for the Boot framework in "Hit the Gym: Accelerating Query Execution to Efficiently Bootstrap Behavior Models for Self-Driving Database Management Systems", appearing in VLDB 2024.

Setup

Read the setup script before running it! It may break your system!

We assume Ubuntu 22.04.

1. Run ./cmudb/setup/dependencies.sh. This sets up:
   • Python3.10 venv
   • Rust
   • Redis
2. Perform standard system setup to reduce noise.
   • e.g., sled experiment checklist

Gather Data

1. Use PGTune to generate a reasonable PostgreSQL configuration for your system and replace the contents of ./cmudb/env/default.pgtune.auto.conf. Otherwise, it defaults to a laptop configuration.
2. Use the virtual environment that was set up above: source ./venv/bin/activate.
3. Invoke ./cmudb/runner/run.sh to run a smoke-test version of all the experiments in the paper.
   • Example changes: SF 1 instead of SF 100, 10 TPC-H seeds instead of 1000 TPC-H seeds, ...
   • Artifacts for each workload (e.g., raw data, models) are stored in the corresponding ./artifact_workload folder.
   • For our example configuration ./cmudb/env/dev8.pgtune.auto.conf, the smoke-test takes around a day to generate data.
4. If you want to run the full set of experiments, run FULL_RUN=true ./cmudb/runner/run.sh.
   • This may take many months on standard server hardware (see ./cmudb/env/dev8.pgtune.auto.conf). That's why this research paper exists!
   • If you have multiple machines with identical specifications, you may modify the script to distribute the work across them. Make sure that for a given Boot configuration and workload, the queries are executed uninterrupted and on the same machine. Otherwise, the Macro-Accelerator may not have the right state.

Plot Results

1. Invoking ./cmudb/runner/run.sh produces an artifact.tgz. Copy and extract this to some folder foo.
2. In the first cell of ./cmudb/runner/analysis.ipynb, set the path ARTIFACTS_TGZ_ROOT to foo.
3. If you ran the full set of experiments, also set SMOKE_TEST = False in the same cell. Otherwise, the plots will be different from the paper as you did not generate the necessary data. For example, the plot scripts will substitute both SF100 and SF10 with SF1.
4. Run the notebook to generate all plots. The default save location is foo/plot_camera_ready.