Effects of Routing Computations in ContentBased Routing Networks

Effects of Routing Computations in Content-Based Routing Networks with Mobile Data Sources Vinod Muthusamy, Milenko Petrovic, Hans-Arno Jacobsen University of Toronto August 30, 2005 Eleventh Annual International Conference on Mobile Computing and Networking (Mobi. Com 2005)

Motivation n Explosion in the number information producers q Blogs, wikis, podcasting, photo sharing n Mobility of users q Cell phones, PDAs, sensors n Mobile information producers q Fixed information producers are increasingly mobile q New types of information producers n n Publish/subscribe data dissemination q Well suited to mobile clients n n SMS, camera phones, location-based services Decoupling, filtering Effects of routing computations on pub/sub network with mobile information producers has not been studied q Causes drastically different results and different conclusions Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 2

Publisher Mobility Scenarios n Journalists with blogs q q n 1 Police patrol car q n Update blogs on location Upload pictures from camera phone Status reports about accidents, traffic, crime Mail delivery q 2 Track delivery status, location, broken parts Aug. 30, 2005 (Mobi. Com ’ 05) Information producer Mobile-To. PSS (University of Toronto) 3

Agenda n Publish/subscribe background q q n Model and distributed protocol Routing operations Publisher mobility q Problem n q Solution n n Invalid assumptions cause excessive state maintenance New protocols to distinguish temporary disconnections Evaluation q Effects of routing computations on protocols Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 4

Publish/Subscribe Model Traffic Reports Publisher Cong estio Fire n Publisher lock Grid Nee d Ba nt ckup e d i Acc Publications Snow Robbe ry Broker Network Notification Subscriptions Subscriber Subscription: Name = “Bob” Report = “Accident” Aug. 30, 2005 (Mobi. Com ’ 05) Subscriber Subscription: Injury = True Location = “Cologne” Mobile-To. PSS (University of Toronto) 5

Distributed Publish/Subscribe n Advertisements flooded q n . . . Create ad tree Subscriptions along reverse ad path q n . . . Create multicast tree Publications along reverse sub path Subscriber Publisher Advertisements Subscriptions Publications Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 6

Generic Pub/Sub Router Operations n Advertisement handling q Insert ad into Ads. Table q Find covering ads q Find intersecting subs Insertion e. g. “severity < 4” Covering n n Subscription handling q Insert sub into Subs. Table q Find covering subs q Find intersecting ads Publication handling q Find matching subs Aug. 30, 2005 (Mobi. Com ’ 05) “severity < 4” is covered by “severity < 6” Y N Intersection “severity < 4” intersects “severity > 2” Mobile-To. PSS (University of Toronto) N Y 7

Modeling Routing Computations n Cannot ignore computations in distributed pub/sub protocols n Classes of algorithms q FAST: tuple-based data (e. g. , attribute-value pairs) n q COMPLEX: tree or graph structured data (e. g. , XML, RDF) n n [Fabret et al. , SIGMOD 2001] [Petrovic et al. , WWW 2005] Based on best reported results under most favorable workloads Operation FAST COMPLEX Matching O(1) O(n) Covering O(matching)* Intersection O(matching)* Insertion O(matching)* *Conservative estimates where no data available (Algorithms usually tuned for matching performance) Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 8

Publisher Mobility Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 9

Publisher Mobility Problem n n Ad and sub trees Moveout: both trees torn down Movein: both trees rebuilt Expensive q q n . . . 1 2 # ad messages > # sub messages No delivery until tree constructed Distinguish temporary disconnections t 1 t 3 At Old Broker t 4 Disconnected t 5 At New Broker moveout t 2 Disconnect (moveout) Aug. 30, 2005 (Mobi. Com ’ 05) Connect Can publish (movein) new events Mobile-To. PSS (University of Toronto) Publisher 10

Publisher Mobility Problem n n Ad and sub trees Moveout: both trees torn down Movein: both trees rebuilt Expensive q q n . . . 1 2 # ad messages > # sub messages No delivery until tree constructed Distinguish temporary disconnections t 1 t 3 At Old Broker t 4 Disconnected t 5 At New Broker movein t 2 Disconnect (moveout) Aug. 30, 2005 (Mobi. Com ’ 05) Connect Can publish (movein) new events Mobile-To. PSS (University of Toronto) Publisher 11

Prefetching Protocol n Exploits knowledge of future mobility patterns Concurrent q Construction at new broker q Teardown at old broker n Tree construction time hidden from user n t 1 t 3 At Old Broker t 4 Disconnected t 5 At New Broker . . . 1 2 moveout t 2 Disconnect (moveout) Aug. 30, 2005 (Mobi. Com ’ 05) Connect Can publish (movein) new events Mobile-To. PSS (University of Toronto) Publisher 12

Prefetching Protocol n Exploits knowledge of future mobility patterns Concurrent q Construction at new broker q Teardown at old broker n Tree construction time hidden from user n t 1 t 3 At Old Broker t 4 Disconnected t 5 At New Broker . . . 1 2 movein t 2 Disconnect (moveout) Aug. 30, 2005 (Mobi. Com ’ 05) Connect Can publish (movein) new events Mobile-To. PSS (University of Toronto) Publisher 13

Proxy Protocol n n Maintain trees from several brokers Advantageous if restricted mobility region t 1 t 3 At Old Broker t 4 Disconnected t 5 At New Broker . . . 1 2 moveout movein Publisher t 2 Disconnect (moveout) Aug. 30, 2005 (Mobi. Com ’ 05) Connect Can publish (movein) new events Mobile-To. PSS (University of Toronto) 14

Delayed Protocol n n n Maintain trees at old broker for some time Allow new tree to graft onto old tree Remove extraneous portions of old tree t 1 t 3 At Old Broker t 4 Disconnected t 5 At New Broker . . . 1 2 moveout movein Publisher t 2 Disconnect (moveout) Aug. 30, 2005 (Mobi. Com ’ 05) Connect Can publish (movein) new events Mobile-To. PSS (University of Toronto) 15

Prefetch-Delayed Protocol n Combine advantages of q . . . 1 2 moveout movein Publisher Prefetching n q . . . Tree construction time hidden from user Delayed n Cheap tree construction cost t 1 t 3 At Old Broker t 4 Disconnected t 5 At New Broker t 2 Disconnect (moveout) Aug. 30, 2005 (Mobi. Com ’ 05) Connect Can publish (movein) new events Mobile-To. PSS (University of Toronto) 16

Evaluation Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 17

Evaluation Setup n Simulation Environment q q n ns-2 network simulator Implemented mobility protocol optimizations Parameters q Topology n n q q q Subscribers: 500 Publishers: 50 Mobility n n n Metropolitan Area Network 4 levels of degree 4 64 leaf brokers • • • 1 64 Static subscribers, mobile publishers Random speeds (5 km/h, 50 km/h, 100 km/h) Metrics q q Tree rebuild load Tree rebuild time Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 18

Routing Computation Model n Based on conservative estimates or best published results of four pub/sub router operations q For COMPLEX: assume n = 100 000 subscriptions at each broker n Ignore other processing delay sources q Network protocol stack, operating system, etc. Operation FAST COMPLEX Matching 2 ms n / 1000 + 40 ms Covering 2 ms n / 1000 + 40 ms Intersection 2 ms n / 1000 + 40 ms Insertion 2 ms n / 1000 + 40 ms n = number of subscriptions Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 19

Publisher Scalability – No Routing Computations n Standard is much worse than Proxy which is worse than Delayed, Prefetch-Delayed q n For both tree reconstruction message load and time Tree reconstruction time seems independent of number of publishers q Incorrect conclusion No algorithm Aug. 30, 2005 (Mobi. Com ’ 05) No algorithm Mobile-To. PSS (University of Toronto) 20

Publisher Scalability – With RC n With FAST algorithm q q n With COMPLEX algorithm q q n Scale: approx. 5 X worse Trend: varies with number of publishers (no longer independent) Standard protocol collapses after 150 publishers 60 s tree rebuilding time with 250 publishers! Routing computations can alter the apparent scalability of protocols q Network is not necessarily the bottleneck FAST algorithm Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) COMPLEX algorithm 21

Implications of Slow Tree Rebuilding n Publications sent during tree rebuilding may not be delivered n 100% delivery with no routing computation q False impression of protocol’s performance Aug. 30, 2005 (Mobi. Com ’ 05) STANDARD protocol Mobile-To. PSS (University of Toronto) 22

Proxy Locality – With RC n n No change in the trends Change in point where Proxy outperforms Prefetch. Delayed q The relative negative impact of overshoot on Proxy increases with more expressive subscription languages No algorithm Aug. 30, 2005 (Mobi. Com ’ 05) FAST algorithm Mobile-To. PSS (University of Toronto) COMPLEX algorithm 23

Conclusions n The publish/subscribe model is well suited to mobile applications q n Routing computations cannot be ignored in pub/sub protocols q q q n No evaluation of mobility with routing computations Not a lower order effect Affect the scale, trend, tradeoff points of results Alter the conclusions of protocols’ performance Future Work q Refine computation models n q Expect greater impact of routing computations Other scenarios: realistic traces, mobile subscribers Aug. 30, 2005 (Mobi. Com ’ 05) Mobile-To. PSS (University of Toronto) 24