Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

20
Distributed Storage Coding for Flexible and Efficient Data Dissemination and Retrieval in Wireless Sensor Networks Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings 1

description

Distributed Storage Coding for Flexible and Efficient Data Dissemination and Retrieval in Wireless Sensor Networks. Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings. Outline. Introduction Related Works Network model Efficient Dissemination of Source Data - PowerPoint PPT Presentation

Transcript of Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Page 1: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Distributed Storage Coding for Flexible andEfficient Data Dissemination and Retrieval in

Wireless Sensor Networks

Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou

IEEE ICC 2010 Proceedings

1

Page 2: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Outline• Introduction

• Related Works

• Network model

• Efficient Dissemination of Source Data

• Proposed Distributed Storage Coding Scheme

• Scheme Evaluation

• Conclusion

2

Page 3: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Introduction• The technique of distributed storage coding has been widely

used in WSN for increasing the data persistence and efficiency of data retrieval.

• Centralized storages such as sink cause performance bottleneck.• Sensed data should be distributed over the whole network.

• How to disseminate data efficiently is a problem.• Many exiting works utilize the random walk to disseminate data.• However, traditional random walk algorithm is inefficient.

• The scheme proposed in this paper ensures the efficient dissemination and flexible recovery of data of interest.• Based on a variant of Metropolis Algorithm• Taking advantage of broadcast transmission.

3

Page 4: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Goal• Improving the efficiency of traditional random walk algorithm.

• Querying storage nodes at the t-th time slot to recover the up-to date data generated by the target source nodes.• each storage node encodes all the source packets received from

the same source node into one packet.

4

Page 5: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Related works• A decentralized coding algorithm based on Reed-Solomon code.

[1]

• ML decoding algorithm makes the decoding process inefficient.

• Taking advantage of the linear coding property of Fountain codes.[2]-[5]

• The drawback of decoding all packets or nothing.

• Addressing partial data recovery problem.[6]-[9]

• Not flexible

5

Page 6: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Related Works[5]

• The probabilistic forwarding tables for random walks are computed by the Metropolis algorithm based on the required steady-state distribution of the random walks, which in turn is derived from the initially assigned RSD.

() the code degree of node i > node j ‘s

• The transmission cost is the product of and .

6

[5] Y. Lin, B. Liang, and B. Li, “Data persistence in large-scale sensor networks with decentralized fountain codes,” in INFOCOM, May 2007

Page 7: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Related Works[5]

7

Step 1 : Degree generation• Choose degree independently from RSD.

Step 2 : Compute steady-state distribution

Step 3 : Compute probabilistic forwarding table• By the Metropolis algorithm

Step 4 : Compute the number of random walkStep 5 : Block disseminationStep 6: Encoding

• Choose d distinct packets from received packets to be encoded.

This paper,

Page 8: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Network model• Random geometric graph, denoted G( n, r).

• k source nodes.

• The sensing time is divided into K time slots.

• Each source node produces one packet per time slot.

• Adopt the same asynchronous time model as that used in [11].

8[11] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, “Gossip algorithms: design, analysis and applications,” in INFOCOM, 2005, pp. 1653–1664.

Page 9: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Efficient Dissemination of Source Data

• Ideal case• No duplicate source packet arrives at the same storage node.• A storage node x should receive d(x) packets from each source

node over K time slot.• n storage nodes will receive packets totally.

• The transmission cost is the product of and .

• Reduce

• employ the broadcast nature of wireless transmission.

• If one packet is buffered on a node for more than one copy, broadcast the packet to all neighbors through one transmission. 9

Page 10: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Efficient Dissemination of Source Data

• To meet the design needs, propose a packet structure

10

• A variant of Metropolis algorithm

Page 11: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Proposed Distributed Storage Coding Scheme

1) Initialization• code degree d(x) • transition probability

2) Data Generation• Generate one data at each time slot t.• Initialize , • according to (1)

3) Data Dissemination4) Data Encoding5) Data Querying and Decoding

11

Page 12: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Data Dissemination• First, the source nodes broadcast source packets to their

neighbor nodes.• Then all the storage nodes use proposed forwarding

algorithm.

12

Page 13: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Example

13

4

1

5

8

9

23

7

611

10

, M = 4L = 75 =23

= 23*1/4 = 5= 23*1/4 = 5= 23*1/4 = 5= 23*1/4 = 5= 3

d=1

d=4

d=2

d=2d=1

d=2

d=3d=3

d=2

d=4

d=1

=5= 5*1/16 = 0= 5*2/16 = 0= 5

=5= 5*1/4 = 1= 4

=5= 5*1/4 = 1= 4

=5= 5*1/4 = 1= 5*1/4 = 1= 3

Page 14: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Data Dissemination

14

Page 15: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Data Encoding• Coded packet consists of

• Encode when a random walk stops at a node with the coded packet of same NID.

• Do not encode the packet

• same PID as the storage node has already encoded

• code degree d(x) is reached

• a new random walk is launched such that .

15

Page 16: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Data Querying and Decoding• Using a mobile collector instead of a sink .

• It can recover all the data in the specific area of interest through querying any storage nodes.• t is the current time slot.• is recovery overhead.

• The collector specifies the NIDs of source nodes deployed in that subarea, and every queried storage node returns a coded packet with specified NID if available.

• Utilize BP decoder to decode all the t source packets of every specified source node.

16

Page 17: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Scheme Evaluation• n = 1000• k = 100• r = 4, avg. neighbors of each node = 4.• 88 time slots, i.e., K =88

• In each time slot, the simulator first disseminates new source packets, and then performs data querying and decoding.

• All results are an average of 50 runs.

17

Page 18: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Recovery Overhead

18

Recovery overhead affected by the length of random walk and the time slot, instead of the fraction of recovered source nodes.

Fig.3(a)Recover 0.2kt through querying storage nodes.

Page 19: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Dissemination Cost

19

Page 20: Ning Cao, Qian Wang, Kui Ren, and Wenjin Lou IEEE ICC 2010 Proceedings

Conclusion• A mobile collector can recover data of any subset of source

nodes at any time via querying any storage nodes.

• Enjoy the broadcast nature of wireless transmission while achieving similar stationary distribution.

• Simulation results show that proposed coding algorithm does not introduce much more overhead to recover data of larger subset of source nodes.

20