Impact of Neighbor Selection on Performance and Resilience of Structured P2P Networks

1
Impact of Neighbor Selection on Performance and
Resilience of Structured P2P Networks

• Byung-Gon Chun, John Kubiatowicz and Ben Y. Zhao

2
Introduction

• Structured p2p overlay networks
• Proximity neighbor selection can lead to
  unbalanced overlay structure
• This paper uses neighbor selection models based
  on network proximity and node capacity
• Drawback increased vulnerability against
  targeted attacks.

3

• Figure 1 Cumulative distribution of node degrees
  of a 205-node Bamboo overlay running on
  PlanetLab. In-degree and outdegree represent the
  number of incoming edges and the number of
  outgoing edges of each overlay node. The graph
  does not include default links (i.e., leafset)
  used for failure tolerance. This is a snapshot
  taken on August 26, 2004.

4
Related Work(1)

• Gummadi et al.(sigcomm03) quantified the impact
  of routing geometry on performance and static
  resilience.
• This paper neighbor selection

5
Related work(2)

• Some works
• ignore node capacity or
• Require static selection supernodes (Brocade) or
• not consider network proximity (Gia in
  sigcomm03, Load balancing in Iptps03)
• This paper
• Combine the heterogeneity of node capacity and
  network proximity

6
Structured overlay construction

• Intelligent selection of neighbors from the set
  of possible neighbor nodes significantly impacts
  the overlays performance, resilience, and load
  balancing properties.
• The neighbor selection problem can be reduced to
  a generalized cost minimization problem.

7
A generalized cost model

• Ideally, optimizing neighbor selection for a node
  means minimizing the sum of the cost from it to
  all other nodes.
• The cost from a node to another consists two
  factors intermediate overlay nodes and overlay
  network links

8
A generalized cost function
9
Cost function for structured networks
10
Idealized cost function vs. practical cost
function

• Idealized require full knowledge of the network
  components, not feasible
• This paper only considers the first hop and
  optimize latency under uniform traffic

11
Four neighbor selection cost functions
12
Simulation setup

• 1. Protocols Tapestry, chord
• 2. For each routing level in Tapestry or each
  finger in Chord, select 32 samples and decides
  the best one to be used.
• 3. 5100 node transit-stub network topologies
  generated using GT-ITM library

13

• 4. chord and Tapestry overlays
• 4096 nodes
• 9 different configurations
• 3 transit-stub topologies per configuration
• 3 overlay node placements per topology

14
Performance

• Node processing delay
• Uniform distribution
• Bimodal distribution
• Fast nodes 100 lookup messages per second
• Slow nodes 1 lookup message per second

15

• Figure 2 Average lookup latency for uniform
  processing delay distribution

16

• Figure 3 Average lookup latency for bimodal
  processing delay
• distribution.

17

• Figure 4 CDF of the number of incoming edges for
  uniform
• processing delay distribution

18
Static resilience

• Metric the proportion of all pairs of live
  endpoints that can still route to each other via
  the overlay after an event
• Nodes capacity distribution is uniform
• For Tapestry three kinds of protocols
• For chord two kinds of protocols

19
static resilience

• Random node failures
• Targeted node attacks
• Analysis of extra redundancy

20
(No Transcript)
21
Figure 6 Chord under random node failures. (a)
Chord varying finger selection on finger table,
(b) Chord varying finger selectionon finger
table and having 4 sequential neighbors.
22
(No Transcript)
23

• Figure 8 Chord nodes under attack. (a) Chord
  varying finger selection on finger table, (b)
  Chord varying finger selection on finger table
  and having 4 sequential neighbors.

24
Conclusion

• A cost model
• Unbalanced structure
• Tradeoff between performance and resilience
• Future work
• Other kinds of geometry
• Dynamic resilience