An event-based http Long-polling Server

Kai Liu, Yilun Cui

Traditionally, the http communication follows the pull model, where
the client initiates the request and server responds it. However, for
web pages that instant between client/server or client/client, the
pull model falls short because the frequent client-initiated
communication will incur a huge overhead.

To overcome the problem, people proposed long (also dubbed ). With the
long polling technology the http server will not terminate a
connection after responded data has been delivered to a client;
instead, the server will keep the connection alive so when new event
occurs the server can send the updated message to the client
immediately.

However this approach requires the server to keep a large amount of
active connections. To efficiently manage the connections we proposed
to add an event-based reverse proxy server in between of client and
backend servers. So when a new client/server session is established,
the reverse will (1) keep the the connection open, and (2) retrieves
resources on behalf of a client from backend servers. Meanwhile the
backend servers can send the updates to the proxy server and the proxy
server will propagate the updates to related active clients.

We plan to benchmark our reverse proxy software with Twitter
stream-like applications which will constantly be publishing new feeds
to a large number of subscribers. We will compare the performance of
the reverse proxy with the traditional opening and closing connection
to evaluate the strengths and weaknesses of our approach.