A Launcher for Distribute ML tasks with BytePS/Horovod

Configuration

Before we start a trial, we need to specify some configuration information in cfg.json. We can add some configuration information to file cfg.json, which is organized in a JSON format. An example of the configuration file cfg.json, here the host field corresponds to a list of IP address of some host machines, and the field visible_device specify the visible devices on each machines.

{
	"host": [
		"xxx.xxx.xxx.xxx",
		"xxx.xxx.xxx.xxx"
		],
	"visible_device": 0,1
}

When we want to read configuration information about "host", we can enter the bash command python3 utils.py --option readcfg_host.

Usage

Use the following command.

./start <option>

Lauch a Trail

--option start: start remote tasks.

Stop a Trail

--option stop:

Check the Status

--option status: list the status of all containers.

Collect Traces

--option collect: collect all traces to local.