Skip to content

redis-performance/vector-db-benchmark

 
 

Repository files navigation

vector-db-benchmark

Screenshot from 2022-08-23 14-10-01

View results

There are various vector search engines available, and each of them may offer a different set of features and efficiency. But how do we measure the performance? There is no clear definition and in a specific case you may worry about a specific thing, while not paying much attention to other aspects. This project is a general framework for benchmarking different engines under the same hardware constraints, so you can choose what works best for you.

Running any benchmark requires choosing an engine, a dataset and defining the scenario against which it should be tested. A specific scenario may assume running the server in a single or distributed mode, a different client implementation and the number of client instances.

Data sets

We have a number of precomputed data sets. All data sets have been pre-split into train/test and include ground truth data for the top-100 nearest neighbors.

Dataset Dimensions Train size Test size Neighbors Distance
LAION-1M: subset of LAION 400M English (image embedings) 512 1,000,000 10,000 100 Angular
LAION-10M: subset of LAION 400M English (image embedings) 512 10,000,000 10,000 100 Angular
LAION-20M: subset of LAION 400M English (image embedings) 512 20,000,000 10,000 100 Angular
LAION-40M: subset of LAION 400M English (image embedings) 512 40,000,000 10,000 100 Angular
LAION-100M: subset of LAION 400M English (image embedings) 512 100,000,000 10,000 100 Angular
LAION-200M: subset of LAION 400M English (image embedings) 512 200,000,000 10,000 100 Angular
LAION-400M: from LAION 400M English (image embedings) 512 400,000,000 10,000 100 Angular

How to run a benchmark?

Benchmarks are implemented in server-client mode, meaning that the server is running in a single machine, and the client is running on another.

Run the server

All engines are served using docker compose. The configuration is in the servers.

To launch the server instance, run the following command:

cd ./engine/servers/<engine-configuration-name>
docker compose up

Containers are expected to expose all necessary ports, so the client can connect to them.

Run the client

Install dependencies:

pip install poetry
poetry install

Run the benchmark:

Usage: run.py [OPTIONS]

  Example: python3 -m run --engines *-m-16-* --datasets glove-*

Options:
  --engines TEXT                  [default: *]
  --datasets TEXT                 [default: *]
  --host TEXT                     [default: localhost]
  --skip-upload / --no-skip-upload
                                  [default: no-skip-upload]
  --install-completion            Install completion for the current shell.
  --show-completion               Show completion for the current shell, to
                                  copy it or customize the installation.
  --help                          Show this message and exit.

Command allows you to specify wildcards for engines and datasets. Results of the benchmarks are stored in the ./results/ directory.

How to update benchmark parameters?

Each engine has a configuration file, which is used to define the parameters for the benchmark. Configuration files are located in the configuration directory.

Each step in the benchmark process is using a dedicated configuration's path:

  • connection_params - passed to the client during the connection phase.
  • collection_params - parameters, used to create the collection, indexing parameters are usually defined here.
  • upload_params - parameters, used to upload the data to the server.
  • search_params - passed to the client during the search phase. Framework allows multiple search configurations for the same experiment run.

Exact values of the parameters are individual for each engine.

How to register a dataset?

Datasets are configured in the datasets/datasets.json file. Framework will automatically download the dataset and store it in the datasets directory.

How to implement a new engine?

There are a few base classes that you can use to implement a new engine.

  • BaseConfigurator - defines methods to create collections, setup indexing parameters.
  • BaseUploader - defines methods to upload the data to the server.
  • BaseSearcher - defines methods to search the data.

See the examples in the clients directory.

Once all the necessary classes are implemented, you can register the engine in the ClientFactory.

About

Framework for benchmarking vector search engines

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 81.2%
  • Shell 13.9%
  • Jupyter Notebook 4.6%
  • Dockerfile 0.3%