Topology Mapping Bo Sheng Sept. 15
Outline Overview Solutions LTM ACE Problems and discussion Conclusion
Introduction Topology mapping Mismatch between overlay and physical infrastructure Topology optimization
Introduction Traffic problem Facts Reasons 95% of any pairs of Gnutella nodes are within 7 hops 50,000 nodes generate 1G/second, 330T/month Reasons Blind flooding Cycles, merge of multiple paths, neighbors exchange Topology problem Multiple times over a physical link
Introduction Perfect match S S Network infrastructure Overlay network
Introduction Mismatch N3 N1 4 5 3 2 S S 2 5 4 N2 Network infrastructure Overlay network
Topology Mismatch Problems Randomly choosing neighbors Logically close, but physically far away S P N1 N2
Topology Mismatch Problems Unnecessary traffic Delayed response Inefficient utilization of bandwidth Only 2%~5% Gnutella connections link nodes within a single AS (autonomous system) More than 40% Gnutella nodes are located within top 10 AS Delayed response Do we need long-distance neighbors?
Topology Mismatch Solutions to traffic problem Selective flooding Topology optimization Avoid cycles Mapping For each message, how many times it is delivered over a single physical link?
Performance Metrics Traffic cost Search scope Response time Overhead
Approaches Location-aware Topology Matching (LTM), INFOCOM 2004 Adaptive Connection Establishment (ACE), ICDCS 2004
LTM Three main operations TTL-2-detector flooding Message format Short Source IP& timestamp Long Source IP& timestamp, TTL1 IP& timestamp d(i,S,v) Link cost IP(S),T(S) S N1 N2 IP(S),T(S) IP(N1),T(N1) d(i,S,1) d(i,S,0)
LTM Three main operations Low productive connection cutting Case1: P receives d(i,S,1) and d(i,S,0) S N P will-cut list
LTM Three main operations Low productive connection cutting Case2: P receives multiple d(i,S,0) S N1 N2 P
LTM Three main operations Low productive connection cutting Case3: P receives one d(i,S,1) and multiple d(i,S,0) S N1 N2 P cut list
LTM Three main operations Source peer probing S N1 P
LTM Step2.case2 S S Step3 N1 N1 N2 P P
LTM Step2.case3 Step2.case2 S S N1 N1 N2 N2 P P Step2.case3
LTM Step3 S S Step2.case1 N1 N1 P P
LTM States Case2 Case1 Case3 Step3
LTM Performance Traffic Search scope Overhead
ACE Step1: Probe link costs with neighbors Build neighbor cost table Exchange neighbors cost table with neighbors
ACE Step2: Create a minimum spanning tree among each peer and its neighbors E E 14 14 4 4 15 G G S S 6 6 20 F F
ACE Step3: Replace neighbors Case1: SH<SG E Case2: GH>SH>SG 14 4 Case3: SH>SG,SH>GH G S 6 H F
ACE Depth of optimization (h-neighbor closure) A B D C E A->B=10 15 B 20 D 8 12 14 C E 7 A->B=10 A->D=15 E->C=7 E->D=14 B->E=8 D->E=14 Total:68
ACE 2-neighbor closure A A D D B B C E C E A->B=10 B->E=8 15 D D B 20 B 8 12 14 C E C E 7 A->B=10 B->E=8 E->C=7 E->D=14 Total:39
Discussion Measurement Link cutting and cycles Link cost is not accurate Link cutting and cycles Heuristic to theoretical support f (Pn,Tn)=?
Conclusion Importance Effectiveness vs. cost Future work