Experiential analysis of Internet AS-level structure
Yangyang Wangadvisor: Prof. Jianping WuNCR, Tsinghua University
outlines
• Motivation• Related work• analysis methods• Data source• Some results• Conclusion
motivation
• ASes are the basic units of Internet• Internet AS-level topology research can help:– Internet topology modeling– Network design planning– architecture design or optimizing for wide-area
distributed system
Related work
• Some related work (not all)– Jellyfish conceptual model (G. Siganos, JCN 2006)• Undirected graph• Degree-based
– BGP routing table tomography from Multiple Vantage Points (L. Subramanian INFOCOM’02)• Directed graph derived from AS_PATH from BGP
monitor
methods• Topology with AS relationship is a directed graph G
– Each vertex is AS– directed edge: connection between ASes – customer-provider, peer-peer, sibling-sibling
Provider
Customer
one directed edge represent a C2P relationship
Peer/Sibling
Peer/Sibling
P2P or S2S relationship is split into two inverse directed edges
methods
• Ranking algorithmsr = 1while (leaves(G) is not empty){
for each v of leaves(G)rank(v) = R;
remove leaves from G, get a new G;r = r + 1;
}for all the rest v in G
rank(v) = r;
• Case 1:• Leaves(G) are vertexes
without customers. It just considers the C2P hierarchy, ignore P2P and S2S.
• Case 2:• Leaves(G) are vertexes
without customers, peers and siblings.
Data source
• CAIDA inferred AS relationship dataset– http://as-rank.caida.org/data/
– Its raw data are from RouteViews project and IRR(Internet Routing Registry)
• Data in 03/03/2008-03/17/2008– as-rel.20080303.a0.01000.txt
as-rel.20080310.a0.01000.txtas-rel.20080317.a0.01000.txt
– Joined them into one data file, choose the inferred relationship with higher occurrence frequency as the final relationship for each link
layershierarchical layers of Case 1
24415 (87%)
2468 (9%)
584 (2%)
176 (1%) 202 (1%)
Outmost layer with very high number of AS
These middle layers include very little number of AS
Innermost layer
layershierarchical layers of Case 2
23767 (85%)
2014 (7%)
308 (1%)
54
1820 (7%)
bigger Innermost layer than Case 1
Less layers than Case 1
Outmost layer with very high number of AS
Other metrics
• We can examine other aspects:– the number of links in each layer itself– the number of cross links and included P2P links
from one layer to the other layers– average degree of the rest graph G after removing
layer n– link connection density of the rest graph G after
removing layer n, and defined as (E(G)* 2) / (V(G)* (V(G) – 1))
other metrics
layer 1 2 3 4 5 6 7 8 9 10 11(core) p2p link(to core) averagedegree
linksdensity
1 783 7621 5898 3146 2225 929 329 331 45 720 22309 830(85.1%) 4.17 0.0001
2 357 1067 527 309 115 43 32 25 36 4755 860(89.9%) 7.9 0.0022
3 114 280 117 39 18 8 8 13 2019 591(89.5%) 12.48 0.0115
4 55 109 20 18 9 1 9 966 417(87.1%) 16.58 0.0332
5 25 29 5 9 1 13 578 289(90.7%) 18.27 0.0565
6 6 10 2 2 5 264 129(97.7%) 18.47 0.0745
7 2 4 2 1 151 83(97.6%) 17.71 0.0784
8 1 2 3 94 66(97.0%) 16.97 0.0782
9 3 6 29 9(88.9%) 16.43 0.0775
10 2 93 62(100%) 16.38 0.0788
11 1617 770(100%) 16.01 0.0797
cross layer links table of Case 1
many of cross links
few of cross links
plenty of cross links approximate increasing
other metrics
layer 1 2 3 4 5 6 7(core) p2p link (to core) averagedegree
linksdensity
1 645 5477 2424 676 79 1 32364 2(100%) 4.17 0.0001
2 226 488 115 10 0 3839 8(100%) 7.95 0.0019
3 30 66 5 0 800 2(100%) 11.00 0.005
4 8 3 0 173 0 11.84 0.0063
5 0 0 9 0 11.99 0.0066
6 0 15 10(100%) 12.00 0.0066
7 10911 4093(100%) 11.99 0.0066
cross layer links table of Case 2
many of cross links
few of cross links
plenty of cross links
degrees vs. layer rank
Case 2Case 1
we could see that vertexes with high degree may be in lower hierarchy rank
conclusion
• conceptual conclusions– there are a small core (less than 10%), large
margin AS (more than 80%), and middle AS band (about 10%)
– remove margin AS, the rest network size is not very small (about 10%) and connections are very complex
– although outer AS have most connections to the core, there are a few connections to middle AS zone, which reduce the regional routing path.
Problems and future work• Problems:– inaccurate AS relationship
• Missing P2P links– incomplete topology data
• Limited BGP monitor• Measurement bias
• Future work– other method of charactering structure pattern– What’s the size of the core as Internet expands– What would look like the topology structure in the
future 10- 20 years according current principles
Thanks !