Repository files navigation

MLOS is a project to enable autotuning for systems.

• Contents
• Overview
• Organization
• Contributing
• Getting Started
  • conda activation
  • Usage Examples
    • mlos-core
    • mlos-bench
    • mlos-viz
• Installation
• See Also
  • Examples
  • Publications

MLOS currently focuses on anoffline tuning approach, though we intend to add online tuning in the future.

To accomplish this, the general flow involves

• Running a workload (i.e., benchmark) against a system (e.g., a database, web server, or key-value store).
• Retrieving the results of that benchmark, and perhaps some other metrics from the system.
• Feed that data to an optimizer (e.g., usingBayesian Optimization or other techniques).
• Obtain a new suggested config to try from the optimizer.
• Apply that configuration to the target system.
• Repeat until either the exploration budget is consumed or the configurations' performance appear to have converged.

Source:LlamaTune: VLDB 2022

For a brief overview of some of the features and capabilities of MLOS, please see the following video:

To do this this repo provides three Python modules, which can be used independently or in combination:

• mlos-bench provides a framework to help automate running benchmarks as described above.

• mlos-viz provides some simple APIs to help automate visualizing the results of benchmark experiments and their trials.

  It provides a simpleplot(experiment_data) API, whereexperiment_data is obtained from themlos_bench.storage module.

• mlos-core provides an abstraction around existing optimization frameworks (e.g.,FLAML,SMAC, etc.)

  It is intended to provide a simple, easy to consume (e.g. viapip), with low dependencies abstraction to

  • describe a space of context, parameters, their ranges, constraints, etc. and result objectives
  • an "optimizer" serviceabstraction (e.g.register() andsuggest()) so we can easily swap out different implementations methods of searching (e.g. random, BO, LLM, etc.)
  • provide some helpers forautomating optimization experiment runner loops and data collection

For these design requirements we intend to reuse as much from existing OSS libraries as possible and layer policies and optimizations specifically geared towards autotuning systems over top.

By providing wrappers we aim to also allow more easily experimenting with replacing underlying optimizer components as new techniques become available or seem to be a better match for certain systems.

SeeCONTRIBUTING.md for details on development environment and contributing.

The development environment for MLOS usesconda anddevcontainers to ease dependency management, but not all these libraries are required for deployment.

For instructions on setting up the development environment please try one of the following options:

• seeCONTRIBUTING.md for details on setting up a local development environment
• launch this repository (or your fork) in acodespace, or
• have a look at one of the autotuning example repositories likesqlite-autotuning to kick the tires in acodespace in your browser immediately :)

1. Create themlos Conda environment.

   conda env create -f conda-envs/mlos.yml

   See theconda-envs/ directory for additional conda environment files, including those used for Windows (e.g.mlos-windows.yml).

   or

   # This will also ensure the environment is update to date using "conda env update -f conda-envs/mlos.yml"make conda-env

   Note: the latter expects a *nix environment.

2. Initialize the shell environment.

   conda activate mlos

For an example of using themlos_core optimizer APIs run theBayesianOptimization.ipynb notebook.

For an example of using themlos_bench tool to run an experiment, see themlos_bench Quickstart README.

Here's a quick summary:

./scripts/generate-azure-credentials-config> global_config_azure.jsonc# run a simple experimentmlos_bench --config ./mlos_bench/mlos_bench/config/cli/azure-redis-1shot.jsonc

See Also:

• mlos_bench/README.md for more details on this example.
• mlos_bench/config for additional configuration details.
• sqlite-autotuning for a complete external example of using MLOS to tunesqlite.

For a simple example of using themlos_viz module to visualize the results of an experiment, see thesqlite-autotuning repository, especially themlos_demo_sqlite_teachers.ipynb notebook.

The MLOS modules are published topypi when new releases are tagged:

• mlos-core
• mlos-bench
• mlos-viz

To install the latest release, simply run:

# this will install just the optimizer component with SMAC support:pip install -U mlos-core[smac]# this will install just the optimizer component with flaml support:pip install -U"mlos-core[flaml]"# this will install just the optimizer component with smac and flaml support:pip install -U"mlos-core[smac,flaml]"# this will install both the flaml optimizer and the experiment runner with azure support:pip install -U"mlos-bench[flaml,azure]"# this will install both the smac optimizer and the experiment runner with ssh support:pip install -U"mlos-bench[smac,ssh]"# this will install the postgres storage backend for mlos-bench# and mlos-viz for visualizing results:pip install -U"mlos-bench[postgres]" mlos-viz

Details on using a local version from git are available inCONTRIBUTING.md.

• API and Examples Documentation:https://microsoft.github.io/MLOS
• Source Code Repository:https://github.com/microsoft/MLOS

• mlos-autotuning-template repository

• sqlite-autotuning

  Working example of tuningsqlite with MLOS.

These can be used as starting points for new autotuning projects outside of the main MLOS repository if you want to keep your tuning experiment configs separate from the MLOS codebase.

Alternatively, we accept PRs to add new examples to the main MLOS repository!Seemlos_bench/config andCONTRIBUTING.md for more details.

• TUNA: Tuning Unstable and Noisy Cloud Applications atEuroSys 2025
• MLOS in Action: Bridging the Gap Between Experimentation and Auto-Tuning in the Cloud atVLDB 2024
• Towards Building Autonomous Data Services on Azure inSIGMOD Companion 2023
• LlamaTune: Sample-efficient DBMS configuration tuning atVLDB 2022
• MLOS: An infrastructure for automated software performance engineering atDEEM 2020