IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.

10, October 2009 191

Autonomous Systems From The View Of Multiple

Vantage Points

Sr.Asst.Prof.G.Srinivasa Rao*, Asso.Prof.G.Venkateswara Rao*, Asso.Prof.M.Venkateswara Rao*,

Asso.Prof.G.Appa Rao**, Asst.Prof.D.kishore Kumar*, Asst.Prof.G.Anuradha**, Asst.Prof.D.Veerabadra
Rao*, Asst.Prof.P.Surya Chandra*, K.Krishna Chaitanya*,T.Anand*

* Dept of Information Technology, GITAM University, Visakhapatnam,

*Dept of computer Science And Engineering, GITAM University, Visakhapatnam,

Abstract distribution network (CDN) that can place replicas of the

The delivery of IP traffic through the internet depends on the web site in data centers hosted by different Ass. The
complex interactions between thousands of autonomous systems company can identify the best locations for its replicas.
(ASs) that exchange routing information using the Border As another example, consider a new regional ISP that
Gateway Protocol (BGP). This paper investigates topological
want to connect to a small number of upstream providers.
structure of the internet in terms of customer _provider and peer-
An accurate view of the AS topology and the relationships
peer relationship between AS's, as manifested in BGP routing
policies. We describe a technique for inferring as relationships can help the ISP determine which ASs would provide the
by exploring partial views of the AS graph available from best connectivity to and from the rest of the internet.
different vantage points. Next we are apply the technique to a In this work, we propose a technique for combining
collection or ten BGP routing tables to infer the relationship data from multiple vantage points in the internet to
between neighboring ASs .Based on these results , we analyze construct a more complete view of the topology and the
the hierarchical structure of the Internet and purpose a five-level AS relationships. Each vantage points offer a partial view
classification of ASs. Our characterization differs from previous of the internet topology as viewed from one source node.
studies by focusing on commercial relationships between ASs
Due to the presence of complex routing polices, these
rather than simply the connectivity between the nodes.
partial views are not necessarily shortest path trees and
Keywords: may, in fact, include cycles. We generate a directed AS
Computer Networks: Autonomous Systems from the view of level graph from each vantage point and assign a rank to
multiple vantagepoints each AS based on its position. Then, each AS is
represented by the vector that contains its rank form each
Introduction: of routing table dumps.
We infer the relationship between two ASs by
TODAY’s internet is divided into more than 10,000 comparing their vectors. Based on these relationships, we
Autonomous systems (ASs) that interact to coordinate the construct a new directed AS graph and examine the AS
delivery of IP traffic. An AS typically falls under the level hierarchy of the internet. We present a five-level
administrative control of single institution, such as classification of ASs with the top most level that consists
university, company or internet service provider of a rich set of peer–peer relationships between 20 so
(ISP).Neighboring ASs use the Border gate way protocol called tire-1 providers.
(BGP) to exchange information about how to reach the The work we describe is novel in three ways. First, we
individual blocks of destination IP address (prefixes). analyze AS path seen from multiple locations to from a
AS relationships and the associated routing polices have more complete views of the graph. Second, rather than
a profound influence on how traffic leaves through the simply combining the data from the various vantage points,
internet. An understanding the structure of the internet in we propose a methodology for exploiting the uniqueness
terms of these relationships facilitates a wide range of of view to infer the relationships between AS pairs. Third
important applications. For example, consider a content we characterize the hierarchy of Ass based on the

Manuscript received October 5, 2009

Manuscript revised October 20, 2009
192 IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.10, October 2009

commercial relationships, rather than simply the in Europe may be a customer of an AS in the US
connectivity of the graph. We evaluate our technique on a for some destinations and a peer for others.
collection of ten BGP routing tables and summarize Router misconfiguration also causes violations in
characteristics of the AS relationships. the export rules. For example, a customer may
mistakenly export advertisements learned from
Problem formulation: one provider to another.

In this section, we present a brief overview of how AS Ranking:

AS relationships affect BGP export polices and formally
define the type of relationship (ToR) problem. Then we Our algorithm assigns a rank to each AS for each
discuss the practical challenges that arise in solving this vantage point. Let X denote the source AS of a particular
problem and validating potential solution. view of the AS graph and let P(X) denote the set of AS
paths seen from X .Since each p ε P(X) consists of a
A. Type of Relationship problem: sequence of notes starting with X, we construct a
directed graph G routed at X from P(X). Let c(Gx)
The relationships between ASs arise from contracts that
define pricing model and the exchange of traffic between denote the set of all vertices in Gx and let levels (Gx) ‫ؿ‬
domains. ASs typically have a provider-customer or peer-
peer relationship.
v(Gx) denote the leaves of the graph. We assign a
In a provider-customer relationship, the customer is ranking rank(u) to each vertex u ε v(Gx) by applying the
typically a smaller AS that pays a larger AS for access to reverse pruning algorithm in figure1 . At each edge the
the rest of the interne. The provider may in turn be a algorithm identifies the leaf node, assigns them a rank
customer of an even larger AS. In a peer-to-peer and remove these nodes (and their incident edges ) from
relationships of the two peers are typically of comparable the graph. In the end of the remaining nodes (if any)
size and find it mutually advantageous to exchange traffic form the connected component of the original graph G\:
between their respective customers. These relationships these nodes are assigned the same (highest) rank.
translate directly into policies for exporting route
advertisements via BGP sessions with neighboring Ass.
G = GX;
a. Exporting to a provider: In exchanging routing r = 1;
information with a provider, an AS can export its while (leaves(G) = φ)
routes learned from other providers or peers. {
for all u ‫ א‬leaves(G)
b. Exporting to a peer: In exchanging routing rank(u) = r;
information with a peer, an AS can export its v = v(G) − leaves(G);
routes and routes of the customers, but cannot r = r + 1;
export routs learned from other providers or peers. G = Gv ;
}
c. Exporting to customers: An AS can export its for all u ‫ א‬v(G)
routes, routes of its customers, but routes learned set rank(u) = r;
from other-providers and peer to its customers.
Fig1: Reverse pruning algorithm on Graph Gx
B. Practical challenges:

Indentifying the commercial relationships

between Ass in challenging in practice. First, Multiple vantage points:
peer-peer relationships are difficult to classify.
As pair may have different relationships for If we continuing the pruning in figure2(a) the
certain blocks of IP address. For example, an AS eventual leaf C will be inferred as a customer of D, even

 
IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.10, October 2009 193

though the two ASs have a pear-pear relationship. relationship may not be visible from same partial views
Identifying the boundary points the uphill and downhill because some ASs may assign to low preference to path
portions of the path is tricky. The structure of partial that traverse in this edge. We use the probalistic
view of the AS graph depends on the position of the AS equivalence rule to find peering edges where the
in the internet hierarchy. In the figure3, the boundary is relationship between two ASs is not visible from many
between C and D (peer-peer relationship) not at E as partial views.
suggested by this partial view in the figure2(a). Now
consider the view from AS B in figure2(b) .This view • Equivalence two ASs i and j are said to be
confirms that A is a customer of D and F is a customer equivalent if e(i,j)>N/2. This rule considers two
of E, however, the graph contradicts the previous view ASs that have the same rank in more than half the
in that d is a customer of C. clearly D and C cannot be vantage points. In these ASs share an edge, they
customers of each other. This contradiction suggests that are likely to be peers.
the two ASs may have a peer-peer relationship.
• Probabilistic equivalence two ASs are probably
1/δ1 δ1 close to 1. We use this rule to infer
peering relationships between ASs when
visibility is poor across the partial views.

2. Inferring provider-customer relationships:

We use the dominance rule to determine if an edge

between two ASs is a provider customer relationship
because one AS tends to have a higher rank than the other
in many of the partial views. Typically, in the graphs from
the vantage point of J or its customers, it is the probable
that rank (j) >rank (i) even if is a provider of j. to avoid an
incorrect inference in such cases, we use the probabilistic
dominance rule.

• Dominance an AS i is said to be dominate AS j if

1(j,j) >= N/2 and 1(j,i)=0. If dominates j, then we
can infer that i is the provider of j, if the two ASs
share an edge.

• Probabilistic dominance If for the high value of

then i probably dominates j, and thus i is a
provider of j. δ0 should be greater thanδ1. We use
the value of 3 for δ0 in our experiments

The orthogonal equivalence and the dominance rules infer

peer-peer and provider customer relationships with a high
degree of confidence. We apply those rules first in our
inference algorithm, followed by the two probabilistic
rules. Those AS relationships which are not infer using
Inference rules of the ToR problem: l ( i,j )
these have the values of max( l ( j,i) ,ll((j,ii,j))) δ1 between
1. Inferring peer-peer relationships: δ0.

We use equivalence rule below to identify the peer-peer

edges that are visible from many views. An AS

 
194 IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.10, October 2009

Experimental Results: Modeling the topology as a directed graph provides a more

precise characterization of the bottom-most level in the AS
This section evaluates our inference techniques on a hierarchy, in the directed graph constructed from the BGP
collection to ten publicly available BGP Routing tables. dumps, 8898 of the 10915 ASs are leaf nodes. The rest of
We classify our relationships between ASs and identify a the graph contains just 18.5% of the ASs.
small number of AS paths that of inconsistent with a
relationship agreement. The most common anomalies 2. Dense core:
seam to steam from resent acquisitions and mergers,
suggest that some AS pairs have a sibling relationship. i). Identifying the dense core:

First we order vertices based on “greedy” notation on

Internet hierarchy:
connectivity, following the heuristic in figure4. Let G
In addition to relationship between AS pairs it is useful represents the directed graph representation of the core let
to identify the position of each AS in the internet p(G) and E(G) represents the vertices and the edges of the
hierarchy. Previous work has classified ASs based on the
node degree: ASs with a large number of neighbors are graph G. let d(x,y) for x ‫ א‬v(G) denote the number of the
placed above ASs with small node degree. However, a
simple degree based approach may not capture the essence
of the tiers in the hierarchy. Instead, we classify ASs
based on the commercial relationships. Typically, a
customer should be at a low level in the hierarchy than edges of the form(x,y) where z ‫א‬Y. connectivity from a
provider(s) we represent the AS topology as directed
graph, where the direction of an edge indicates the type of
relationship between two ASs. In our graph a provider-
node to a given set of nodes refer to the number of
customer relationship between A and B is represented by a
directed edges from the node to any of the nodes in the set.
directed edge from A to B and a peering relationship
Assume that k of the N nodes are already ordered, for each
between A and B is represented by two directed edges,
of the remaining N-k nodes, we determine the connectivity
one from A to B and other from B to A. in such a graph
to the k node and the pick the node with maximum
representation has also been independently proposed. An
connectivity as the(k+1). When multiple nodes have the
important difference between our approaches is the
same connectivity, with higher outer degree. In figure4,
procedure used for determining the internet hierarchy. The
pos(x) denotes the position of the node X in the final
work in maps the internet topology in to a strict hierarchy
ordering.
based on provider-customer edges while our classification
also uses the distribution of peering links as identify the
compute z ‫ א‬v(G) with maximum out-
top levels of the hierarchy.
degree;
1. Customer and small regional ISPs X = {z};
pos(z) = 1; r = 1;
Customers are the easiest class of ASs that can be classified while (X = v(G)) {
from these directed graph structure of AS topology. compute y ‫ א‬v(G) − X with max
Customers are those sub networks which are origin and d(y, X)
sinks of traffic and which do not carry any transit traffic. (selecting the y with the max
out-degree)
From the very definition of the direction of the edges in our X = X ‫{ ׫‬y};
graph we can infer the customer ASs to be the leaves of the maxindegree(r) = d(y, X);
directed graph. In a directed graph, a leaf is a node with r = r + 1;
out- degree 0. Since and undirected graph makes no pos(y) = r;
distinction between out-degree and in-degree, customers }
with multiple providers would have a degree more than 1 Fig:4
and hence would not appear as leaves of the graph.

 
IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.10, October 2009 195

many of the top providers in Europe and Asia to be

ii). Transit core: present in our transit core.

After removing the dense core, we notice the presence of iii). Outer core:
other large national providers and hosting companies that
We classify all the remaining ASs in the core as the outer
have peering relationships with many of the ASs in the
core. The members of the outer core typically represent
dense core. To identify these ASs, we define the notation
regional ISPs which have a few customer ASs and a few
of a transit core. Nodes in the transit core peer with each
peering relationships with other such regional ISPs. The
other and with ASs in the dense core, but they do not tend
outer core consists of 897 ASs that have 29 peering
to peer with many out ASs. In our directed graph
sessions with ASs in the dense core and 145 peering
representation, these peering links are essentially the
sessions with ASs in the transit core. We observed that
incoming directed edges from vertices outside this set to
many members of our outer core are regional ISPs.
vertices within the set. We define such a set of edges to be
in-way cutoff graph indicated by the given set of vertices. iv). Summary:
Using the property, we define the transit core as a smallest
set of ASs containing the dense core which indicates a Table1, shows summarized number of ASs at each level
weak in-way cut. That is one having small number of in the hierarchy dense core(level 0), transit core(level 1),
edges compared to the total number of ASs in the transit outer core(level 2), small regional ISPs(level 3) and
core. customer (level 4). Table2, summarizes the connectivity
between various levels in the AS hierarchy . each number
in the table is the total number of edges from level 0 to
1. Identifying the transit core: given X ‫ ؿ‬v(G), let
level 1. The table shows several key properties of the
internet topology.

cutin(X) denote the set of all edges of the form (y,z). • The ASs in dense coer are veru well connected.
We define a cut X of the vertex set v(G) to be a weak
cut in |cutin(X)| >X/2. The problem of finding weak • As we move from the dense core towards
cuts in the graph is NP-complete and no good customer, the inter-level and intra-level
approximation algorithms are known for that problem. connectivity graphs significantly.
Given the transit core is a superset of the dense core
and that the dense core is derived by the greedy • The large number of customers ASs have their providers
ordering, we apply some ordering to find the transit distributed across levels is not strictly hierarchal.
core as was used to find the dense core. A natural way
• The number of edges within the outer core is less
of using this ordering to find the transit core is to find
than the total number of vertices in the outer core.
the smallest value of k such that |cutin(Xk)|<k/2.
This indicates the presence of multiple
Surprisingly we find that the value of k at which |cut in
disconnected graphs of ASs in the outer core;
(Xk)|< k/2 also satisfy the property that conn(k+1)=1.
Level # of Ass
This means that no two edges in the cutin(Xk) have the
same source. A weak cut also means that more than Dense core(0) 20
50% of the ASs in Xk do not have any peering
relationship with any of the ASs in v(G)-Xk. Hence by Transit core(1) 129
this definition, Xk should intended contain all the
transit providers. Outer core(2) 897

2. Properties of the transit core: Applying the in-way Small regional ISPs 971
cut algorithm to our graph, we discover a transit core
consisting of 129 ASs. These 129 ASs have 183 Customers 8898
peering links with the ASs in the dense core. We found

 
196 IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.10, October 2009

ASs in the different groups communicates via The complete structure of the internet is unknown and
ASs in the dense core and transit core. difficult if not impossible to obtain. Our approach is
comprised of many heuristics with certain limitations.
Distribution of ASs in the Hierarchy(Table1)
• We draw our inferences based on only ten
Interconnectivity Across levels(Table2) vantage points available. Ideally we would have a
The graph in figure5, explores the relationship between
node degree and the levels in the hierarchy. We define Level 0 1 2 3 4
node degree as number of neighboring ASs without regard
0 312 626 1091 958 6732
to the relationships. The graph plats the cumulative
distribution of node degree as a logarithmic scale. In 1 183 850 1413 665 3373
general level 0 and 1 ASs have high degree, and level3
and level4 ASs tends to have low degree. However, this is 2 29 145 1600 543 3752
not universally true. Some customers at level 4 have a
large number of upstream providers, and some ASs in the 3 0 0 0 212 2400
dense core at level 0 have a relatively small number of
larger collection of routing tables from more
neighbours. A hierarchy based solely on degree
diverse vantage points, including small customers.
distribution would not be able to make this distribution.
• We treat the route views routing tables as a view
from a single AS.

Future work:
We plan to extract a separate view for each AS
participating in the route views project. Multiple ASs
may fall under the administrative control of a single
institution, due to historical artifacts and market forces.
We plan to extend our methodology to incorporate more
complex routing policies that are captured by the
Fig5: Cumulative distribution of AS degree by level traditional customer-provider and peer-peer relationship.

Despite of these limitations, we have shown that our

approach provides a detailed view of the Internet
Summary: topology in terms of the relationships between ASs.
The relationships between ASs has a significant impact
References:
on the flow of traffic through the internet. Our work
makes two important contribution toward understanding [1] John W. Stewart, BGP4: Inter-Domain Routing in the
the structure of the internet in terms of these relationships: Internet, Addison-Wesley, 1998.
[2] Sam Halabi and Danny McPherson, Internet Routing
• An algorithm for inferring AS relationships from Architectures, Cisco Press, second edition, 2001.
[3] Geoff Huston,“Interconnection, peering, and
partial views of the AS graph from different settlements,” in Proc.INET, June 1999.
vantage points. [4] Alaettinoglu, “Scalable router configuration for the
Internet,” in Proc. IEEE IC3N, October 1996.
• A mechanism for dividing the Internet hierarchy [5] B. Krishnamurthy and J. Wang, “On network-aware
clustering of Web clients,” in Proc. ACM SIGCOMM,
into levels based on AS relationships and node
August/September 2000.
connectivity. [6] R. Govindan and H. Tangmunarunkit, “Heuristics for
Internet map discovery,” in Proc. IEEE INFOCOM,
March 2000.

 
IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.10, October 2009 197

[7] H. Tangmunarunkit, R. Govindan, S. Jamin, S. Shenker, Project in 2007 and published 8 papers in various National and
and W. Willinger,“Network topologies, power laws, and International Conferences and Journals.
hierarchy,” Tech. Rep. 01-746, Computer Science,
University of Southern California, June 2001.
Mr. M.Venkateswara Rao., M.Sc,
[8] “CAIDA Web Site,” http://www.caida.org. M.Phil (Ph.D)l, and M.Tech in
[9] R. Govindan and A. Reddy, “An analysis of interdomain Computer Science Engineering from
topology and route stability,” Proc. IEEE INFOCOM, Andhra University. Published 4 papers
April 1997.
in various National and International
[10] L. Gao, “On inferring automonous system relationships
in the Internet,”IEEE/ACM Trans. Networking, vol. 9, Conferences and Journals.
no. 6, December 2001.
[11] “University of Oregon RouteViews
project,”http://www.routeviews.org/.
[12] “BGP tables from the University of Oregon RouteViews
MR. G.Appa Rao., M.Tech., M.B.A.,
Project,”http://moat.nlanr.net/AS/Data/.
[13] Michalis Faloutsos, Petros Faloutsos, and Christos (Ph.D)., in computer science and
Faloutsos, “On power- law relationships of the Internet Engineering form Andhra Universiy. He
topology,” in Proc. ACM SIGCOMM, August/September has Submitted Ph.D thesis in M.U., Over
1999, pp. 251–262. 12 Years of teaching experience with
[14] Jin, Q. Chen, and S. Jamin, “Inet: Internet topology GITAM University, handled courses for
generator,” Tech. Rep. CSE-TR-433-00, U. Michigan, B.Tech, M.Tech. Research areas include
September 2000.
Data Mining and AI. Published 8papers in various National and
[15] Z. Ge, D. Figueiredo, S. Jaiwal, and L. Gao, “On the
hierarchical structure of the logical Internet graph,” in Proc. International Conferences and Journals.
SPIE ITCOM, August 2001.
[16] Lixin Gao, Timothy G. Griffin, and Jennifer Rexford, Mr. D.Kishore Kumar M.Tech in
“Inherently safe backup routing with BGP,” in Proc. IEEE computer science and engineering. Over 4
INFOCOM, April 2001. Years of teaching experience with
[17] L. Gao and J. Rexford, “Stable Internet routing without GITAM University, handled courses for
global coordination,” IEEE/ACM Trans. Networking, vol.
B.Tech.
9, no. 6, pp. 681–692, December 2001.
[18] H. Chang, R. Govindan, S. Jamin, S. Shenker, and W.
Willinger, “On inferring AS-level connectivity from BGP
routing tables,” 2001.
[19] “RADB Whois Server,” whois.radb.net. Mr. K.Krishna Chaitanya M.Sc.,
M.Tech(IT) from GITAM University
[20] “IXC Buys Three ISP’s,”

Mr. G.Srinivasa Rao, M.Tech,(Ph.D).,

Sr.Asst.Professor. He has Submitted Ph.D
thesis in M.U., Over 7 Years of teaching
experience with GITAM University,
handled courses for B.Tech, M.Tech.
Research areas include Computer
Smt. S.Anuradha., M.Tech., (Ph.D). ,in computer science and
Networks And Data Communications.
engineering. Over 4 Years of teaching experience with GITAM
published 4 papers in various National and International
University, handled courses for B.Tech and M.Tech .
Conferences and Journals.
Mr. D.Veerabadra Rao., M.Tech., in computer science and
Mr. G.Venkateswara Rao Received M.Sc,
engineering. Over 4 Years of teaching experience with GITAM
M.Phil, and M.Tech in Computer Science
University, handled courses for B.Tech and M.Tech .
Engineering from Andhra University
during 1993, 1996 and 1999 respectively.
Mr. P.Surya Chandra., M.Tech in computer science and
He is working as faculty member in
engineering. Over 3 Years of teaching experience with GITAM
GITAM University in the Dept. of
University, handled courses for B.Tech and M.Tech.
Information Technology from last 10 years.
His research interest includes Mobile
Mr. T.Anand ., M.Tech(IT) from GITAM University
computing and Networks. He was awarded UGC Major Research