CrypTen Cifar10 Distributed Demo with CUDA

This project shows how to run CrypTen Cifar10 demo on distributed servers with CUDA enanbled.

Environment preparation

Install CUDA environment, pytorch, and make sure they are working normally.
Install CrypTen using source (recommended).
Glone this project, and change into the project directory.

Data preparation

Manually download cifar10 python dataset cifar-10-python.tar.gz and put it in data foler, or you can ignore the step and let the program download the dataset automatically each time it runs.

Test locally on single server with two Nvidia GPUs

Assume you have two parties, start the first party process by

# rank 0
python launch_distributed.py  --evaluate \
    --world_size 2 \
    --rank 0 \
    --master_address 127.0.0.1 \
    --master_port 12345 \
    --gpu_id 0 \
    --model-location model/model_best.pth.tar \
    --resume  \
    --batch-size 50 \
    --print-freq 1 \
    --skip-plaintext \
    --distributed

Start the second party process by

# rank 1
python launch_distributed.py  --evaluate \
    --world_size 2 \
    --rank 1 \
    --master_address 127.0.0.1 \
    --master_port 12345 \
    --gpu_id 1 \
    --model-location model/model_best.pth.tar \
    --resume  \
    --batch-size 50 \
    --print-freq 1 \
    --skip-plaintext \
    --distributed

Note the master_address, master_port, and gpu_id can be changed but must be available.

Test distributedly on two servers with an Nvidia GPU

Ensure the two servers are connected with on the same network.
Assume you have two parties, and the master node IP address is 192.168.0.1.
Start the first party process on the master node by

# rank 0
python launch_distributed.py  --evaluate \
    --world_size 2 \
    --rank 0 \
    --master_address 192.168.0.1 \
    --master_port 12345 \
    --gpu_id 0 \
    --model-location model/model_best.pth.tar \
    --resume  \
    --batch-size 50 \
    --print-freq 1 \
    --skip-plaintext \
    --distributed

Start the second party process the other node by