Realcouse2.0: From Fully Connected Topology to Multiple Overlapping Stars

We present the second version of Realcourse. Realcourse is a cooperative Internet video publishing and delivery system used by more than thirty Chinese universities to share course videos. The design of Realcourse bears the following assumptions 1) the system is supposed to consists of hundreds of unreliable but controllable servers where temporal failures are common, permanent failures are rare, 2) videos are concurrently uploaded to local servers but their metadata are broadcasted to all servers, 3) users are connected to and served by local servers but should be able to watch all the videos in the system in a seamless and efficient fashion. Realcourse has gone through two versions. In Realcourse1.0, all servers were treated equal and fully connected by messaging channels. In order to maintain a consistent global directory among servers, each server informs all others of local update operations through the message channels. The fully connected server topology poses a great challenge to system management with respect to failures. To solve this problem, Realcourse2.0 adopts a server topology of multiple overlapping stars by 1) replacing direct broadcast with relayed-broadcast, in which message is first sent to a relay server and then forwarded to all other servers, 2) partitioning the global directory into sub-trees, each sub-tree is then fully replicated to a group of selected servers while groups may overlap with each other. This paper explains why such a change is necessary and how it was done. Realcourse has been operational since December 2003. Today, more than 5000 videos are served. We constantly observe about 10000 unique user IP addresses each day.