Predictive Publications and Subscriptions

Rapidly moving objects (for example, players) in the game need to publish updates at the game's frame rate for other nodes to discover them quickly. Given the high frame rates of these games ($\sim$10 frames per second), this results in a significant amount of network traffic. Instead, objects send out predictive publications - these are events which indicate where the object may be expected to be found in the next few frames - using a range of x,y & z coordinates. These publications may result in some false matches, but they allow the publisher to re-publish at a much lower rate. We show that the resulting overhead due to false matches is much less than the bandwidth savings achieved. Predictive publications are implemented as publications containing constraints - thus, they look like subscriptions, except carry publication semantics. A predictive publication matches a subscription if their ranges overlap.

It turns out that sending predictive subscriptions is also important, although for a different reason. Players in a game like Quake II tend to move at a very fast rate and so their viewable area changes rapidly. Hence, some amount of interest ``pre-fetching'' is required to hide the latencies involved in object discovery. This is achieved by sending subscriptions which are slightly larger than the currently visible bounding box.