This script is adapted from https://github.com/nwstephens/triton-xgboost. It demonstrates an error when using the perf_analyzer between two containers in a docker network. The workaround is to use `--network host` instead of `--network=tritonnet`. For more information on the implications using the host network, see: https://docs.docker.com/netwo…

perf_analyzer_issue.sh

# Create docker network
sudo docker network create tritonnet

# Create shared volume for model_repository
sudo docker volume create volume1

# Pull and run Triton container
sudo docker pull nvcr.io/nvidia/tritonserver:22.03-py3
sudo docker run --gpus=all -d -p 8000:8000 -p 8001:8001 -p 8002:8002 --network=tritonnet \
    -v /var/lib/docker/volumes/volume1/_data/model_repository:/models \
    --name tritonserver nvcr.io/nvidia/tritonserver:22.03-py3 tritonserver --model-repository=/models

# Pull and run PyTorch container
sudo docker pull nvcr.io/nvidia/pytorch:22.03-py3
sudo docker run --gpus=all -t -d -p 8888:8888 --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --network=tritonnet \
    --mount source=volume1,destination=/workspace/volume1 --name pytorch nvcr.io/nvidia/pytorch:22.03-py3

# Performance analyzer throws an error from inside the PyTorch container
sudo docker exec -it pytorch /bin/bash
git clone https://github.com/nwstephens/triton-xgboost.git
cp -R triton-xgboost/data/pre-built/ volume1/model_repository/.
pip install tritonclient[all]
apt update
apt install libb64-0d
perf_analyzer -m pre-built --percentile=95 #error: failed to get model metadata: HTTP client failed: Couldn't connect to server
exit

# Performance analyzer should be successful from the host machine
perf_analyzer -m pre-built --percentile=95