Docker and horovod command collection

docker_horovod_collection.sh

# build images from Dockerfile
docker build -t horovod:latest horovod-docker-gpu

# get list images
docker images

# get list container
docker ps

# remove container
docker rm [container name]

#remove images
docker rmi [image nama:tag]

# open container
docker exec -it [container-id] bash

# save images to tar
# example : nvidia-docker save -o horovod.tar horovod:latest
nvidia-docker save -o [name of file.tar] [name of images:tag]

# load docker tar to images
docker load -i horovod.tar

# run docker with nvidia runtime
docker run --runtime=nvidia -it horovod:latest

# run docker with specific ssh network
docker run --runtime=nvidia -it --network=host -v /home/kvgn-beta/.ssh:/root/.ssh horovod:latest

# open connection
/usr/sbin/sshd -p 12345; sleep infinity

# run docker and open ssh connection for specific ports
docker run --runtime=nvidia -it --network=host -v /home/kvgn-beta/.ssh:/root/.ssh horovod:latest \
    bash -c "/usr/sbin/sshd -p 12345; sleep infinity"
    
# run ini distributed machine, 4 machines with 1 GPU each
# --start-timeout important for running multiple machine, because loading distributed model and optimizer quite long
# for the first time
horovodrun -np 4 --start-timeout 300 \
  -H localhost:1,192.168.99.2:1,192.168.99.12:1,192.168.99.4:1 -p 12345 \
  python pytorch_mnist.py

# Other helping commands
# add user 
adduser [name of username]

# change network name, from eno2 to eno1
sudo /sbin/ip link set eno2 down
sudo /sbin/ip link set eno2 name eno1
sudo /sbin/ip link set eno1 up

# split file into 500MB each
split -b 500M dynamic.tar

# join files
cat x* > dynamic.tar