The server doesn't "request" the client, rather the client initialises the conversation by opening an HTML 5 WebSocket. That socket then stays open, and the server can push data back to the client at any time.
You can do this in HTML 4 using "long polling", where a client makes a request and the response is also kept open for a longer period. The problem here is that the client cannot send another request on the same socket, so needs to hold two sockets open to the server - one to send request, the other to received responses. This is only bad if you have many many clients attached to the server and are already on the limit of how many connections your server can handle.
In any case, you want to configure the server to use "non-blocking" (NIO) connections, so that it doesn't need one thread per client, because handling anything more than a few thousand concurrent threads seems to bring the server to a stop. Using a non-blocking solution, you can theoretically have 50 thousand or more connections per server instance, although how useful that will be is questionable, because of the amount of requests it can process quickly with that many connections. If every client makes a request per second, you will be handling 50 requests per millisecond, which is unlikely to give you enough time to do anything resonable with say a database.
See the following links:
http://blog.maxant.co.uk/pebble/2011/06/21/1308690720000.html
http://blog.maxant.co.uk/pebble/2011/05/22/1306092969466.html
http://blog.maxant.co.uk/pebble/2011/03/05/1299360960000.html
