High performance and high scalability are fundamental requirements for a streaming/push server, that has to support many concurrent connections and deliver data to the client in the shortest possible time.

Lightstreamer Server is a real-time system implemented in Java. In order to guarantee the maximum level of performance, the Server has been developed as a stand-alone process which does not rely on an underlying web server or application server. In this way it is possible to have direct control over the TCP/IP layer of the system and over the threading model of the application. True push solutions need to maintain at least one open TCP socket for each connected client. Some web/application servers have been extended to act as streaming engines, but their traditional architecture based on the “one-thread-per-connection” model makes scalability pretty impossible to tens of thousands of users. Lightstreamer adopts a staged event-driven architecture instead of a thread-based one, making it possible to decouple the number of connections that the Server can sustain from the number of threads that are employed. The Java NIO classes are fully exploited to implement this model, through asynchronous I/O, for both HTTP and HTTPS connections.

Lightstreamer Server process can scale well over multiple CPUs. In case of CPU saturation, Lightstreamer offers a “graceful degradation”, i.e. the system does not hang abruptly but progressively decreases its performance while guaranteeing service continuity. The system can scale horizontally too by adding more boxes beyond a common web load balancing appliance in order to form a cluster.

Capacity planning depends on many variables. The most important are the inbound and outbound event throughputs, message sizes, maximum latencies. For these reasons the best way to decide the required hardware is to run specific simulations for the necessary scenarios. A specific Load Test Toolkit can be provided to run the simulations. Just to give a general idea about Lightstreamer’s scalability, when data push happens at low frequency (around 1 update per second for each user), 10,000 concurrent users can be handled by a single CPU.