MOH server is commanded to generate a stream to the IP address of the endpoint receiving the stream. The endpoint is told to listen to the stream, which includes various messages depending on the protocol and the type of endpoint. When the call is to be taken off hold, the MOH server is told to stop streaming, and the held party is connected to the appropriate endpoint.
MOH server is always streaming the multicast stream. As such, CallManager doesn't need to tell the MOH server to do anything. The endpoint is simply told to listen to the multicast stream.