Topology Aware Overlay Networks

1
Topology Aware Overlay Networks

• Junghee Han, David Watson, and Farnam Jahanian
• Department of Electrical Engineering and Computer
  Science University of Michigan
• 2005 Infocom

2
Roadmap

• Introduction
• Background
• Measurement methodology node placement
  heuristics
• Single vs. Multi-Hop Overlay Routing
• Evaluation
• Conclusion

3
Introduction

• Current underlying routing protocols
• Overlay network
• Loss and failure correlation between overlay
  paths at the underlying IP layer

4
Introduction (cont,d)

• Three contributions
• Design a overlay network to maximize path
  independence
• Single hop overlay path provided the same path
  diversity as multi-hop
• Validation based on real Internet failures

5
Background To demonstrate path diversity issue
How many overlay links shared at least one router
at the IP layer with li
50 overlay nodes from planetlab
lith

• Proposes a topology-aware overlay to max path
  diversity for better availability and performance

6
Topology-Aware Node Placement

• Assumption
• Consider these routers as possible places to
  deploy overlay nodes
• Full deployment of overlay nodes on all routers
  at all ISPs
• Impractical and high cost in deployment
• topology-aware node placement
• Measurement-based analysis to identify the most
  cost-effective depth and breadth of node
  placement
• Depth
• Compare routers in each ISP
• Breadth
• Compare different ISPs

7
A. Measurement Methodology

• Definition
• Direct path
• Indirect path
• Metrics to determine the quality of each overlay
  node ni
• Path diversity
• of shared routes between Direct path and
  Indirect path
• Latency
• RTT difference between Direct path and Indirect
  path

8
To gather the direct and indirect path
information

• Apply this procedure to the 100(100100) pairs
  of source and destination

9
B. Placement of overlay nodes inside an ISP
network
Fig. 4. Comparing path diversity of different
overlay nodes within the same ISP

• Leftmost line
• Dynamically select the optimal overlay node
• Others statically selected
• No single node provide the best path diversity gt
  more than one!

10

• Q
• How many routers are needed to provide optimal or
  nearly optimal diversity?
• Which routers within the same ISP should we
  choose?
• Clustering-based heuristic
• Two overlay nodes fall into the same cluster when
  their path diversity patterns are similar
• Use correlation to identify path diversity
  patterns of all src and dst

11
Path diversity

• S src hosts
• D dst hosts
• N of s/d pairs (i.e.,100200)
• Is,d of overlapping routers between
• the indirect overlay path through overlay
• node i and direct path from s to d

12

• Latency overhead (Performance)
• In contrast to path diversity, overlay nodes
  inside the same ISP show very different patterns
  of latency.
• Selection of overlay nodes becomes more critical
  with respect to latency rather than path
  diversity.

Fig. 6. Comparing latency of different overlay
nodes within the same ISP
13
Apply the same clustering method except that we
use RTT instead of path diversity to calculate
the similarity between overlay nodes
14
Combine path diversity and latency

• Assign a tuple (cp,cl) to each router,
• Cp is the cluster number with respect to path
  diversity
• Cl is the cluster number with respect to path
  latency
• Then, we perform the iterative search to select
  one router from each cluster

15
C. Choosing a set of ISP network

• Two data set
• 11 ISPs
• Using K3 cluster-based
• methods

• Statically
• Overall path diversity provided by the individual
  ISP is almost the same
• Dynamically
• Carefully choose more than one ISP in order to
  provide better path diversity.

16
How many ISPs?
Fraction of pairs with no overlapping between
the direct Internet path and indirect overlay
paths

• Careful selection of ISPs ? 4 is important
• gt 4-ISP selection, the choice of ISPs does not
  matter much

• ISPs ? 3 provides only marginal gains
• 4 routers per each ISP is most cost-effective
  regardless of the number of selected ISPs

17
How the selection of an ISP affects latency

• The leftmost line
• optimal case where we dynamically select the best
  ISP depending on the source and destination.
• The choice of ISP does not matter with respect to
  latency.
• Select routers inside the ISP becomes critical

18
Single vs. Multi-Hop Overlay Routing

• RON
• Most path outages were avoided by detouring
  through only a single overlay node
• This paper complement their studies by providing
  measurement-based verification
• single-hop overlay routing performs as well as
  multi-hop routing in terms of both availability
  and performance

19
Compare Diversity

• D1 and D2, described in Section III-A.
• DPs,d direct path
• SP is,d single-hop overlay path through node i
• MPs,d optimal overlay path

Single-hop overlay path almost as good as
optimal multi-hop overlay path does
20
Compare Latency

• Single overlay nodes
• One between the source and an overlay node, and
  the other between the overlay node and a
  destination
• Multiple overlay nodes
• calculate the shortest roundtrip time for all
  overlay path

• Single-hop overlay paths better than direct
  Internet paths ,as well as the shortest multi-hop
  paths

21
Evaluation

• Statically choose ISPs and routers based on the
  clustering mechanisms proposed
• Adopt the proposed single-hop routing
• Include logs of failure events in the real world

22
Select ISP(I)

• No single ISP does provide the best path
  diversity for every source and destination pair.
• 2-ISP deployment as good as full deployment(87)
• A carefully selected subset of ISPs can perform
  as well as full deployment throughout different
  time periods

23
Select Routers in an ISP(II)

• Evaluation results
• Sprint and Qwest
• The period of April 12-15
• Statically pick 6 (or 3) routers based on the
  measurements in Section III-B.
• 3-cluster deployment provides almost the same
  degree of recovery as full deployment

24
Conclusion

• The proposed approach is able to react and
  recover from about 87 of path outages
• Existing overlay networks were unable to avoid
  about 50 of path outages
• Our proposed heuristics for choosing ISPs and
  overlay nodes provide almost the same degree of
  resilience as full deployment