1

Detecting and Reducing Partition Nodes in Limited-routing-hop

Overlay Networks

Zhenhua Li and Guihai Chen

State Key Laboratory for Novel Software TechnologyNanjing University, Nanjing, P. R. China

lizhenhua@dislab.nju.edu.cn, gchen@nju.edu.cn

2

Background

• Overlay networks - base infrastructures of many Internet applications

• Limited routing hops - routing one hop in the overlay network is much more

expensive than that in the underlying network.

- flooding or flooding-based routing mechanism

- so has a limit called TTL

3

Motivation

• Overlay partition - seriously degrade the system performance

• Existence of topologically-critical nodes - some nodes’ failure will cause overlay partition with much

higher possibility than others

4

Related work (1)

• Proactive avoid and Event driven - using a centralized server to direct nodes’ join and leave

- but the server becomes a single point of failure

• Proactive avoid and Periodical detect - CAM: actively detect cut nodes and then neutralize them

into normal nodes

- but cut nodes are not applicable to limited-routing-hop overlay networks

5

Related work (2)

• Reactive recover and Event driven - ring partition detect and repair on Pastry and SkipNet

- but they can only be used on ring topology

• Reactive recover and Periodical detect - cross-check method: ask other nodes to do random queri

es and compare their results with its own

- but it has much randomness and uncertainty of detection

6

• The concept of partition node - topologically-critical nodes of limited-routing-hop overlay

networks

• Partition node detection and reduction - a distributed proactive method to detect partition nodes

- reduce partition nodes by changing them to normal nodes

- greatly enhance the connectivity and fault tolerance of overlay networks

Our proposed ideas

7

Outline

• Partiton node concept

• Partition node detection

• Partition node reduction

• Performance evaluation

8

Partition node concept (1)

• Cut node vs. partition node.

- (a) (b) C is a cut node because when C fails, the overlay network is partitioned;

- (c) (d) C is a partition node because when C fails, the overlay network is not partitioned, but C’s neighbors 1, 3, 5, 7 can no longer find each other.

9

Partition node concept (2)

Definition 1 (Locatability) In a limited-routing-hop overlay network, node A could locate node B only if A can find B by

sending routing messages. It is denoted by A→B.

Definition 2 (Reachability) In a limited-routing-hop overlay network, node A could reach node C if A can locate C, or A can locate some node B and B can locate C. It is denoted by A→→C.

10

Partition node concept (3)

• Example: Node 1 can only locate nodes 2, 3, 4, and can reach node 5, 6, 7, but cannot reach node 8.

11

Partition node concept (4)

• Definition 3 (Partition Node) Node C is a partition node if C’s neighbor set would be partitioned into two or more unreachable subsets S1, S2, . . . , Sn (n≥2) when C fails.

• Example:

12

Partition node detection (1)

4 steps: • Initialize detection (0)(1)• Probe reachability (2a)(2b)• Partition subsets (3)• Make decision (4)

13

Partition node detection (2)

14

Partition node reduction (1)

• Add edges to reduce partition nodes - choose an appropriate delegate node Ni from each subse

t Si,

- and then connects all the delegate nodes in some way.

- In order to improve the system’s fault tolerance, we try to make every node’s degree above a constant lower bound as much as possible.

15

Partition node reduction (2)

• Linear chain connection vs. Chordal ring connection

- more edges, but much more resilience

16

Partition node reduction (3)

• Remove edges to limit node degree - the new edges added to reduce a partition node cannot

be removed; - remove the edge whose corresponding node has the hi

ghest load factor.

• Total cost of partition node detection and reduction

- n: tatal number of nodes, t: TTL, c: average node degree

- total cost is ))(),(min( 2cnOncO t

17

Performance evaluation (1)

• Partition nodes’ significance to overlay topology.

18

Performance evaluation (2)

• Effectiveness of our method

19

Performance evaluation (3)

• Fault tolerance improvement

20

The End

Thanks!