The preferred method of using Docker is to just to download the publicly released pre-built docker images. In addition, the enclosed Dockerfile can be built to produce an image capable of running GPUSimilarity.
- nvidia-docker (regular docker will not expose GPUs)
- Newest versions of nvidia drivers
- Recommended for production: systemd / nginx
Both releases and master are always built and uploaded publicly already. You can access them via: For master:
docker pull klorton/gpusimilarity:latest
For releases:
docker pull klorton/gpusimilarity:v1.0
The underlying requirements will change very little, so after the initial build/pull new docker layers will only require a few megabytes for gpusimilarity changes.
From the docker directory:
docker build -t gpusim:internal .
docker run -v /PATH/TO/SMIGZ_DIR:/data -it gpusim python3 \
/gpusimilarity/bld/python/gpusim_createdb.py /data/INPUT.smi.gz \
/data/OUTPUT.fsim
This will run single threaded. If you're building huge databases you should enter the container interactively with the data directory mounted, and follow the regular instructions from the parent ReadMe to run multithreaded.
/bin/nvidia-docker run --net=host -v /path/to/fsim/files:/mnt/fsim -it \
klorton/gpusimilarity:latest python3 /gpusimilarity/bld/python/gpusim_server.py \
/mnt/fsim/1.fsim --port 8080 --http_interface
Once the server says "Ready for Searches", you should be able to test it by either using a graphical web browser or links to access "http://localhost:8080"
See the provided service file to be placed in /etc/systemd/system/
- The
--http_interfaceoption is only for testing never use it in production as it hasn't been hardened at all - You should use a systemd script to start/stop the service, including restarts
- Never expose the http from this service directly, use nginx or equivalent with certificates to expose via ssl/https.