If I understand correctly, Hadoop or BigData applicatins are highly I/O bound. So for performance processing you would definitely prefer many little physical machines. Apart from that setting up Hadoop on VMWare should bring no questions that would not occur on physical installations.
Depending on your performance needs and your three virtual machines you would probably go with a dedicated master and two slave nodes. The master should be running NameNode and ResourceManager, the slaves DataNode and NodeManager. Other nodes in the Hadoop setup are optional.