Network Topology Measurement Yang Chen CS 8803 1

advertisement
Network Topology Measurement
Yang Chen
CS 8803
1
Outline
•
•
•
•
Big Picture
ISP Topology Measurement
Statistical Results
Problems & Solutions
2
Heuristics for Internet Map
Discovery
R. Govindan and H. Tangmunarunkit
INFOCOM 2000
3
Why do we need the topology?
• Understand the macroscopic properties of
the Internet physical structure
• Network management
• Topology-aware algorithms
• Simulation and topology generation tools
4
On-going efforts
• CAIDA Skitter
• Router View
5
Fundamental: traceroute
Prober sends packets with successively increased TTL.
A router responds with ICMP time exceeded when the
probe is with TTL=1
6
Fundamental: traceroute
Geographic info can help on building up the topology.
* Data
from http://www.linkwan.com/vr/
7
Fundamental: Tree => map
1)Source routing
2)Multiple Vantage points
8
Address scan space
•
•
•
•
BGP tables
Route Table
Database
Informed Random Address Probing
– A response from some IP address is considered as a
sign that some prefix P of A must contain addressable
nodes;
– If P is an addressable prefix, the neighboring prefixes
of P are also considered as possibly addressable.
(128.8/16 and 128.10/16 are neighbors of 128.9/16)
9
Some results
• 150,000 interfaces and nearly 200,000 links
• Findings related to source route
– Simulation demonstrated that In relatively sparse
random networks, a few source route capable nodes
(< 5%) are sufficient to discover 90% of the links. In
fact, there are 8% routers support source route.
– Source route discovered links do not skew the
qualitative conclusion on the network statistics.
10
For example: degree distribution
Similar observation on hop-pair distribution
11
Measuring ISP Topologies
with Rocketfuel
N.Spring, R. Mahajan and D. Wetherall
ACM Sigcomm 2002
12
ISP network infrastructure
Access
Router
Access
Router
Backbone
Router
Backbone Link
13
ISP topology measurement
• An old story: the blinds and the elephant
ISP
Traceroute
14
ISP topology measurement
Traceroute
Server
15
Focusing on one ISP – Directed probing
Network
* 192.9.9.0
*
*
*
*
*>
*
*
*
blackrose.org
Verio
Sprint
MCI
LINX
CERFnet
IIJ
PIPEX
IAGnet
* BGP
Next Hop
204.212.44.128
205.238.48.3
144.228.240.93
204.70.4.89
194.68.130.254
134.24.127.3
202.232.1.8
158.43.133.48
131.103.20.49
(Ann Arbor)
(MAE-WEST)
(Stockton)
(San Francisco)
(London)
(San Diego)
(Japan)
(London)
(Chicago)
M/LP/Weight Path
0 234 266 237 3561 701 90 i
0 2914 1 90 i
0 1239 701 90 i
0 3561 1 90 i
0 5459 5413 1 90 i
0 1740 701 90 i
0 2497 701 90 i
0 1849 702 701 90 i
0 1225 2548 1 90 i
204.212.44.128 through AS234
205.238.48.3 through AS2914
144.228.240.93 through AS1239
204.70.4.89 through AS3561
194.68.130.254 through AS5459
134.24.127.3 through AS1740
202.232.1.8 through AS2497
158.43.133.48 through AS1849
131.103.20.49 through AS1225
table source: RouteView project
16
Focusing on one ISP – Directed probing
• Traceroutes to dependent prefixes: All traceroutes to these
prefixes from any vantage point should transit the ISP.
Dependent prefixes can be readily identified from the BGP
table. All AS-paths for the prefix would contain the number of
the AS being mapped.
• Traceroutes from insiders: We call a traceroute server located
in a dependent prefix an insider. Traceroutes from insiders to
any prefix should transit the ISP.
• Traceroutes that are likely to transit the ISP based on some
AS-path are called up/down traces.
17
Path/Query reduction
Share Ingress
Share egress
Same next-hop
AS number
18
Impacts of directed probing
1) Fraction of useful but pruned traces from 0.1 to 7%
2) Unnecessary traces around 6% over all the ISPs
*Comparison
based on Skitter data
19
Impacts of ingress reduction
Overall, ingress reduction
keeps only 12% of the traces
chosen by directed probing.
The number of vantage points that share an ingress by rank
20
Impacts of egress reduction
Overall, egress reduction
keeps only 18% of the
Dependent prefix traces
chosen by directed probing.
The number of dependent prefixes that share an egress by rank
21
Impacts of next-hop reduction
Overall, Next-hop AS reduction
Reduces the number of traces
to 5% of those chosen by
directed probing.
22
POP sizes analysis
23
Power Law
• Complementary cumulative distribution function
(CCDF) P(X>x)
• Pareto Distribution
 x
P( X  x )  
 xmin



k
P( X  x )  x
k
• Power Law
ln( y )  C ln( x )
24
Router degree distribution
25
Peering structure
26
Difficulties in topology discovery
• Shared media
• Backup links
• Router Identification and annotation
• Alias resolution
• Completeness Validation
Currently, none of them is completely solved!
27
POP hierarchy
Naming convention, DNS information and neighbor inferring
28
Backbone topology
AT&T
Level 3
29
Alias Problem
OR
30
Alias: is it a big deal?
31
Alias resolution
• Send a packet with unreachable port to
certain interfaces which are possible alias.
The corresponding ICMP port unreachable
response will contain the source address.
• IP identifier
32
Completeness validation
• Comparison with Router Views
• Comparison with Skitter
• IP address space
– Search prefixes of ISP’s address space for
additional IP addresses
• Validation with ISPs
– Is “Good” enough?
33
In Search of Path Diversity
in ISP Networks
P. Teixeira, K. Marzullo, S. Savage
and G. M. Voelker
IMC 2003
34
Real metric instead of counting links
• Path diversity
– Metric that reflects the number of routes
available between two points in the network
• An extreme example
35
Real topology speaks
Inter-PoP Path diversity in
the Sprint Network
Inter-PoP Path diversity inferred
by Rocketfuel
36
Take a closer look
37
Inaccuracy introduced during
probing
• Lack of vantage points
– How many points are sufficient?
• Incomplete traceroutes
– What can we do if ISP turns off traceroute
functionality?
• Changes in the path of a probe
• Incorrect DNS record
38
Inaccuracy from processing probed
links
• Alias Resolution
• Adding reverse links
Missed and added links in Rocketfuel PoP topology relative to
the number of links in the Sprintreal topology
39
Questions?
40
Download