REEF currently has a local runtime(reef-runtime-local) for dev/test. It has served its purpose well. However, some users expressed interests in doing this in a distributed fashion.
So, it'd be good to develop a ssh-based standalone distributed runtime. Being ssh-based means that there's less code for us to write and easy for users to deploy. Being standalone means that users don't need to set up separate resource managers(e.g., YARN, Mesos) to run REEF jobs on multiple nodes.
- scp required file to remote nodes and launch them through ssh
- Tasks are submitted through ssh connection and run as foreground, enabling driver to receive return status
- Evaluator logs are saved temporarily on $HOME/REEF_SSH_RUNTIME of each node, which will be sent to driver after task is finished