Flare (Federated Learning AcceleRation Engine) is a distributed machine learning system for Federated Learning. We build Flare based on FATE and deploy it using Docker. Flare provides a high-performance distributed FL computation model and a friendly programming model. The main technologies include dynamic dataflow execution, heterogeneous secret computing accelerator and inter-party communication scheduler.
In this section, we will introduce how to use Docker to deploy flare quickly.
- Docker 18+
- Add user into docker group
Step 1: Change file name Rename flare/fate/confs-* to your party number, such as confs-10000 Step 2: Create folders Transfer path to flare/fate/confs-10000 using cd ./fate/confs-10000 Run: sh create_dir.sh Step 3: Edit confs files Local ip means the ip address of the machine; Docker ip means the ip address of docker network.
- confs/fate_flow/conf/service_conf: change [servings][hosts] to your local ip.
- confs/fate_flow/conf/pulsar_route_table.yaml: Configure routing table.
- confs/fate_flow/conf/rabbitmq_route_table.yaml: Configure routing table.
- confs/nginx/route_table.yaml: Configure routing table.
Step 4: Edit shell files The ports in shell files mean local ports. You must make sure that all ports are not occupied and the ip address of docker network is not occupied as well. For all shell files (such as create_container.sh), you need to edit variables in the front of the shell files, such as USER, BASE_DIR, Party and so on.
We provide three types of container sets:
- standalone: Run flare just in one machine.
- cluster master: Run flare in a cluster and this machine is cluster manager.
- cluster worker: Run flare in a cluster and this machine is cluster worker.