Created by: jacquesqiao

now the configuration for distributed paddle is:

 --pservers=host1,host2 --port=1234

It have some problems:

1. we can only start one trainer on one machine because all trainer share the same port for communication.
2. If the port on one of all the machines is used by other program, the whole job will fail.

Can we modify it to the format like tensorflow, for example:

--cluster_spec='worker|host1:port1,ps|host2:port2'

so that we can start multi trainer on one machine.