1. Improved Algorithms for Network Topology Discovery
Benoit Donnet
joint work with Timur Friedman & Mark Crovella
PAM Boston

2. Context Network measurement
Internet topology discovery using distributed traceroute monitors
IP interface level
Existing tools:
Skitter (CAIDA)
TTM (RIPE NCC)
AMP (NLANR)
DIMES (Tel Aviv U.)

3. Scaling Problem More monitors means more load on
network resources
destinations
Classical approaches either
stay small (skitter, TTM, AMP)
trace slowly (DIMES)
Can we trace more efficiently?

4. Contributions Efficient topology discovery algorithm [Sigmetrics2005]
Doubletree
Reduce communication overhead
Bloom filters
Reduce load on destinations
Capping
Clustering

5. Doubletree - Basics Cooperative algorithm
Goal: avoiding paths already explored
Exploit tree-like structure of routes in the internet
from a monitor to a set of destinations
Backward probing (first suggested by Govindan et al.)
from a set of monitors to a destination
Forward probing and monitor coordination

6. Doubletree: Probing scheme
Two probing schemes:
Backwards
Forwards
Stop sets = {(interface, root)}
Local Stop Set: B = {interface}
Global Stop Set: F = {(interface, destination)}
shared by monitors
Doubletree starts probing at some hop h from the monitor

7. Doubletree - Example

8. Bloom filters - Motivations
Wide deployment of Doubletree difficult
High communication cost due to global stop set sharing

9. Bloom filters - Basics Purpose: reduce communication cost
Lossy summary of a set based on
bit vector
multiple hash functions
Possibility of false positives
We vary
the vector size (from 1 to 27,017,990)
the number of hash functions (from 1 to 5)

10. Bloom filters – false positives

11. Bloom filters - Coverage
Links
Nodes

12. Bloom filters - Redundancy
Destinations
Routers

13. Capping/Clustering - Problem
Load on destinations
Risk of looking like a distributed denial of service attack

14. Capping/Clustering - Problem

15. Capping/Clustering - Basics
Purpose: decrease the load on destinations
Impose a limit on the number of monitors targeting a destination:
Capping: explicit limit
Clustering: group the monitors for each destination

16. Capping

17. Clustering

18. Capping/Clustering - Coverage
Links
Nodes

19. Capping/Clustering - Redundancy
Destinations
Routers

20. Future Work Clustering based on topological information
IP level topology discovery guided by BGP
Doubletree implementation

21. Conclusion Communication cost reduced with Bloom filters
compression factor: 17.3
Load on destinations reduced with
Capping
Clustering

22. Thank you for your attention
Questions?