search_benchmark

Qdrant search speed benchmark

Data

Based on https://github.com/erikbern/ann-benchmarks#data-sets

GloVe Angular 100: http://ann-benchmarks.com/glove-100-angular.hdf5

Usage

Download data

bash get_data.sh

Install dependencies

pip install -r requirements.txt

Run Qdrant instance

docker-compose up

Upload search data

python -m benchmark.upload_data

Uploaded data available for search immediately, but building of HNSW index may take quite some time.

Run search benchmark

python -m benchmark.search

Example Results

Qdrant params:

• Version: v0.3.0
• Num parallel searchers: 4

For GloVe Angular 100 dataset:

• num_vectors = 1,183,514
• dim = 100
• metric = cosine

Used HNSW index params:

• M = 16
• efConstruct = 100
• ef = 100

Search with 1 threads
avg precision = 0.876
total time = 94.418 sec
time per query = 0.0094 sec
query latency = 0.0034 sec

Search with 4 threads
avg precision = 0.876
total time = 28.030 sec
time per query = 0.0028 sec
query latency = 0.0087 sec