PeertoPeer Overlay Networks in an EventBased Middleware Peter
Peer-to-Peer Overlay Networks in an Event-Based Middleware Peter R Pietzuch and Jean Bacon Firstname. Lastname@cl. cam. ac. uk DEBS’ 03, San Diego, CA, USA, June 2003
Overlay Broker Networks • Distributed pub/sub systems – Mapping of brokers to physical nodes – Specification of overlay topology • Efficiency, reliability, manageability, … Today Ó Static neighbour lists – Difficult at deployment time – Require global view – Depend on physical network topology Ó Hierarchical topologies – Hard to maintain global properties – No redundancy Tomorrow E Self-managing overlays – Add more brokers on demand – Not only useful for large-scale E Adaptive overlay networks – Reflect current network situation – Lead to more efficient event dissemination E Evaluation 1
Overview • Overlay Broker Networks • Peer-to-Peer Techniques • Hermes – Type- and Attribute-Based Routing • Simulational Evaluation – Routing Efficiency – Space Efficiency and Distribution – Message Complexity • Conclusions 2
Peer-to-Peer Techniques • Distributed hash tables (Pastry, CAN, Chord, …) – – Overlay network of nodes with unique ids Hash operation from key to nodeid route (msg, key) Scalable and efficient Locality properties • Advantages of P 2 P for publish/subscribe – Higher abstraction for building pub/sub systems – Content-based routing algorithm deals with hash keys – P 2 P overlay handles neighbouring set for event brokers • Easy to manage • Efficient routing • Dynamic mapping • Fault-tolerance 3
Hermes • Hermes, an event-based middleware – P 2 P overlay network – Evaluation of its efficiency vs. Siena-like approach – Type- and attribute-based publish/subscribe • Event Clients • Event Brokers P S P B S B P R B P S B S P B B S S P P P – Local Broker, Rendezvous Node • Rendezvous Nodes – Set up on a per type basis – Hash of event type name gives key for DHT – Ensure that advertisements and subscriptions join in the network 4
Type- and Attribute-Based Routing Advertisement Messages – Routed towards RN by publishers – Create entries in advertisement routing tables along the way Subscription Messages – Routed towards RN by subscribers – Follow the reverse path of advertisements – Create entries in subscription routing tables along the way Publication Messages – Follow the reverse path of subscriptions – Get filtered along the way 5
Simulational Evaluation • Evaluation of content-based pub/sub in simulator – Large-scale deployment for experiments difficult – Realistic network topologies and model for simulation • E. g. notification latency, hop count, routing cost, … – Scale reflects corporate deployment (102 event brokers) • Keep number of subscribers small if routing unaffected DSSim Pan Cov. Adv • Discrete event simulator • Transit stub network model • Visualisation plug-ins • Pastry-like routing • Siena-like pub/sub • Static set of neighbours • Acyclic topology Hermes • Pub/sub 6
Routing Efficiency • Overlay networks used in experiments – Hermes: Sequential addition to closest existing broker w. r. t latency – Cov. Adv: Pre-computed minimum spanning tree Latency per notification (500 brokers; single subscriber per broker) • Quality of the overlay • Decreases tree is more populated • Cov. Adv (closest broker) has worst latency • Cov. Adv (min span) is optimal • Hermes is in between 7
Space Efficiency and Distribution Routing table sizes • Hermes does not broadcast advertisements • Hermes has slightly less subscription state due to better routing • Converge to same value as tables become full Routing table distribution • Cov. Adv: Majority of broker has 20 routing table entries • Hermes: No broker has more than 15; some have none 8
Message Complexity Message numbers (Advs, Subs, Pubs) (100 event types) • Hermes sends fewer publications than Cov. Adv due to its better routing • Hermes sends more subscriptions than Cov. Adv due to RNs • Number of advertisements stays constant 9
Conclusions • Self-managing & adaptive overlay networks are needed – Distributed Hash Tables are helpful • Evaluation through simulation – Contrast different kinds of pub/sub approaches • Peer-to-peer routing with RNs is efficient • Future Work – Fault-tolerance – Dynamic network environments 10
Thank You Any Questions? 11
- Slides: 12