Evaluating P4-based Virtual Delivery Trees for Content-based Publish/Subscribe

With content-based publish/subscribe, publishers issue notifications that are matched against subscriptions issued by subscribers to which only the matching notifications are delivered in turn. Due to the expressiveness of content-based filtering, each notification requires an individual delivery tree. This fact and because subscriptions potentially change over time make it hard to preinstall delivery trees into the switch infrastructure. In this paper, we present novel notification distribution strategies implemented in P4 that install virtual trees (VTs) in the switch infrastructure. For deriving VTs, we use topological information as well as pub/sub-specific information such as the distribution of publishers and subscribers and their interests. A greedy algorithm constructs a notification-specific delivery tree at runtime by combining VTs and encoding additional routing information into the notification headers to extent, connect, and prune VTs. We present a detailed evaluation investigating the performance of the strategies in real-world topologies and under changing subscriptions. The results confirm that our strategies construct efficient distribution trees requiring only little additional routing information in the notification headers. They also show that the size of this routing information grows when subscriptions change and that the increase depends on to what extend the respective strategy tailors the delivery trees to the active subscriptions.