Multicast Protocols for Publish Subscribe Systems Ashwin R

Multicast Protocols for Publish Subscribe Systems Ashwin R. Bharambe Hemamalini Manickavasagam

Outline n n n What is publish-subscribe ? Multicast in publish-subscribe systems Small Group Multicast Evaluation Methodology Results Conclusions

What is publish-subscribe ? Publications Subscription n Publishers produce events or publications Subscribers register their interests via subscriptions Network performs routing such that n n Publications “meet” subscriptions Publications get delivered to appropriate subscribers

Why do we care ? n n De-coupling between senders & receivers Advantages n n Permits heterogeneous components Interoperability Ideal for evolving software Fits many applications n n Stock tickers, instant messengers, etc. Multiplayer games!

Architectures n n n Centralized Rendezvous point based Core components n n n Routing of publications and subscriptions Matching Delivery of publications

Efficient publication delivery n Publication needs to be delivered to some subset of subscribers n n Problem n n n Multicast! Could be any subset Exponential multicast groups! How about the most naïve multicast strategy ?

Small Group Multicast (SGM) n Idea n n Send list of recipients along with every packet! High overhead if list is large, but – More efficient than maintaining multicast group, if list is usually small Use Narada overlay as the backbone

SGM+ n n n SGM might send data to uninterested people Use Narada mesh to get a rough sense of direction Use recursive unicast starting from the closest guy

Evaluation methodology n n Internet games as example of a publish subscribe system Workload n n Quake-II – real life traces hard to find Introduced BOTs into Quake-II n n Supposed to be human-like Visibility “box” used as subscription

Simulation Infrastructure n ns-2 all the way Ported Narada simulator to ns-2 Implemented SGM, SGM+ over Narada

Results (1/2) n SGM overhead prop. to delivery group size n n Hovers around 4. 5 Very high fraction of packets are routed through nodes who are not interested!

Results (2/2) n n Scaled with respect to Unicast As group size increases, SGM+ performs better n Publication delivery group is usually small!

Conclusions n SGM n n Gives good resource usage Low overhead since delivery group sizes in games are typically small Well suited for games! Need workloads for characterizing other publish-subscribe systems