After a cluster has been launched there is no way to access the roles running on particular instances. It would be useful to find the workers in a Hadoop cluster, so you could operate on them (e.g. restarting the daemons on them).
How this is implemented is probably different for different clouds. E.g. on EC2 it could be accomplished using tags, on other platforms a local file might be sufficient, and a blobstore might be another way.