Cycle Detection in PublishSubscribe Overlay Networks Reza Sherafat

Cycle Detection in Publish/Subscribe Overlay Networks Reza Sherafat Alex Cheung Prof. Cristiana Amza ECE 1747 – Course Project University of Toronto 1

Overview of Publish/Subscribe Systems 1. Advertise Publisher 3. Publish Subscriber Network consists of - Clients: 2. Subscribe Publisher, subscriber - Routing brokers Overlay may be cyclic, non-cyclic 2

Cyclic Overlay Networks Supporting cycles is good: • • Load balancing Hard to avoid Failure recovery Routing through cycles is challenging: • • Duplicates waste bandwidth Infinite loops 1 Advertisement Tree 2 S S S Duplicate Messages 2 • 3

Outline • • Routing in cyclic overlay Cycle detection algorithm Implementation Future work 4

Routing Through Cycles Preparation: • • • Detect cycles Let brokers know that they are on cycles Keep cycle information updated Routing: • • Route advertisements in a cycle-free manner Route subscriptions in a cycle-free manner • • Keep track of brokers downstream Route publications in a cycle-free manner • Use cycle information and source routing to exploit cycles 5

Cycles • Large cycles are bad candidates for re-routing • • Alternative paths may be too long We focus on cycles smaller than a configurable size s. • • Better choices for our purposes Easier to detect Cycles detection algorithm, detects such small cycles. 6

Cycle Detection Algorithm We want to detect all cycles smaller than s+1 • Based on Asynchronous Breath-First Search • • Starting from a Heart broker propagate special messages Append broker’s id to the message and send it to neighbors When a duplicate is received, there is a cycle Compare broker id lists to find members Inform all members of the cycle Delays in message delivery are problem some • 7

Cycle Detection Algorithm Compare Propagate b 1, b 1 b 5 b 1, b 2, b 3, b 4, b 5 b 1 s=5 L=4 Ad 1: b 1, b 5 Ad 1: b 1, b 2, b 3, b 4, b 5 Ad 1: b 1, b 2, b 3, b 4 8

Cycle Detection Algorithm • By limiting the size of cycles we lower the overhead • • … … Problem when the “small” cycle is far away from the Heart • • Broker id list will be at most s-1 Heart Because broker id list is too short Hearts must not be far away • Hearts roughly partition the network into local areas A small cycle is not discovered We have to limit the depth d that messages are propagated. 9

Cycle Detection Algorithm Distance from Heart • Roughly speaking we propagate Heart’s messages to double the cycle size (2 s) in order to detect whether a broker within distance s (local) is on a cycle. • • • Hence s brokers in each direction away have a local broker’s id s-1 is sufficient for broker id lists Zones: • • Full discovery zone Partial discovery zone Full Discovery zone (local) Partial discovery zone 10

Cycle Detection Algorithm • Scalability • • • Heart messages are propagated limited # of hops away Broker id lists have at most s-1 ids Each Heart locally detects cycles • We need to place hearts side by side (Future work) 11

Implementation • • Based on pub/sub semantics Message class LOOP_DETECTOR • • Loop detection module • • • Command attribute: control, path, loop An add-on to brokers Process LOOP_DETECTOR publications using separate queues Extensions to the pub/sub semantics • • Source routing Distance-aware 12

Implementation • Path information is part of publication’s payload Brokers subscribe to LOOP_DETECTOR messages Received messages are put into the loopdetector’s queue • Major new classes: • • Loop. Detector. Manager 13

Future Work • Heart placement We need to place hearts side by side • Implement routing algorithm With load balancing • Implement fault recovery algorithm By using alternative paths formed in cycles 14