Data Collection Infrastructure for Location- Location-Unaware Sensor - PowerPoint PPT Presentation

Data Collection Infrastructure for Location- Location-Unaware Sensor Networks Distributed coding protocols for data storage Silvija Kokalj-Filipovic Predrag Spasojevic Roy Yates

Talk OutLine • Data Collection from a Location-Unaware Wireless Sensor Network – Network nodes self-organize into a web-like infrastructure of routes – Network data is encoded and stored along circular infrastructure routes using a distributed coding protocol – A Mobile Data Collector arrives to a random point of the network perimeter – Connects to the closest node of each circular route and collects encoded data from the nodes within its immediate neighborhood – Up-front collection from the neighborhood combined with polling distant nodes selectively to collect symbols which unlock the decoding process is an energy- efficient solution that allows for full decoding – The data collection is completed when the collector decodes all network data

Sensor Network Example location-unaware sensor nodes randomly scattered in a plane

Sensor Network Example Isotropic wireless propagation location-unaware sensor nodes randomly scattered in a plane

Data Dissemination source � advertising along source spokes increases the likelihood of information discovery � avoiding flooding-based data publishing no redundant transmissions (broadcast storm)

Simulated Dissemination Scenario 350 300 250 200 150 100 50 0 −50 0 50 100 150 200 250 300 350

Infrastructure Isometric Routes building Data Collector modeling R 2 R 1 � coding for distributed data storage � decoding strategies for data collection

Isometric Networks network partitioned into sub-networks that are customized to handle network storage task according to the number of associated sources 2 = πλ − E [ k ] r ( 2 i 1 ) i s • Light Isometric Networks • Heavy Isometric Networks • infrastructure developed as a side effect of search for specific data items R 2 R – use for storage of network data, through network- 1 network-coding based methods • inspired by the current work on network coding fo coding for storage in WSN R i = ir Data Collector wants a snapshot of network data

Current Work in Network Coding for Data Storage in WSN Two basic approaches: – Decentralized erasure codes (1) • encode k symbols into codewords of length n, which can be decoded fro m any subset of k symbols within the codeword • Decoding complexity: O(k 3 ) – Decentralized fountain codes (2) • potentially infinitely many codewords (linear combinations of k data blocks); can be decoded from any k independent combinations • Decoding complexity: almost linear in k • Abstract (1) or overly expensive (2) random routing techniques • We propose structured approach to decrease cost (1) Dimakis, Prabhakaran, Ramchandran . Decentralized erasure codes for distributed networked storage ( ‘05/6) (2) Liu, Liang, Li. Data persistence in large-scale sensor networks with decentralized fountain codes ( ‘07)

K. Ramchandran Making dense sensor networks smarter using randomized in-network processing NSF workshop “Future Directions in networked sensing” May 2006 Decentralized Erasure Codes G f 1 X 1 +f 2 X 2 f 1 ⎡ ⎤ f 0 f 0 1 4 f 2 ⎢ ⎥ X 1 = [ Y Y Y Y ] [ X X X ] f f 0 0 f 3 X 2 ⎢ ⎥ 1 2 3 4 1 2 3 2 3 f 3 ⎢ ⎥ ⎣ ⎦ 0 0 f f 5 6 X 2 f 4 f 4 X 1 +f 5 X 3 Want matrix as sparse as possible f 5 X 3 (decreases dissemination cost) f 6 X 3 f 6 k data nodes n storage nodes Now assume only storage nodes 1-3 are queried. To reconstruct it suffices to have G to be full rank

How to build a code if your data is not in one place? Basic Coding Approach: Random Linear Coding Node i holds a codeword of degree d equal to the number of non-zero entries in this column 2 1 i 5 n 1 2 1 1 0 0 0 0 0 0 0 0 0 5 G= T xG= y i n-2 0 0 0 0 0 k n-1 k n Decentralized Fountain: x imposes a probability distribution on codeword degrees Both approaches: a certain number of packet replicas to be randomly diffused from independent independent sources and stored an random nodes (matrix rows)

Isometric Networks network coding selected according to the number of associated sources R i = ir • Light Isometric Networks: – Random Linear Codes • Decoding complexity for L light networks: ( ) 3 ∑ = ∑ L L 3 < 3 ∑ R L k k k 2 R = = i i i 1 i 1 i i 1 1 • Heavy Isometric Networks: – Decentralized Fountain Codes • Decoding complexity for H heavy networks: ∑ = H k i 1 i

Dissemination and Storage Relaying and Overhearing/Combining relays Storage squad: set of nodes in the range of relay � Sources associate with one of the closest relays � Relays disseminate (mix data) � Squad nodes overhear, combine and store data Let us assume that - there is one source per relay (i.e. per squad) - number of sources (relays) n larger than squad size h

Storage and Dissemination Graphs � high energy cost of data dissemination π 2 ir = ≈ π n 4 i i r s Mixing over circular graph with n i nodes, nodes, each of degree 2 Mixing time i 2 /2 Mixing time O( i 2 ) n i 1 2 3 Super-Squads: sets of adjacent squads • Circular dissemination with network coding in the context of “wireless multicast multicast advantage” • Apply network coding for storage in squad nodes that overhear 2 relays

Storage in Isometric Networks Super-squad h squad nodes 0 0 0 0 0 0 0 0 0 h=O( r s 2 ), h<n n sources 0 0 0 0 0 hn storage nodes Random Matrix created by Storage Protocol MDC collects from SUPER-SQUADS Storage Protocol controls degree distribution of codewords large number of sources: storage with linear decoding complexity needed COLLECTION Assumption: a mobile data collector (MDC) will establish connection with a random relay Goal: to have all data of the isometric network available in the vicinity of selected relay

Collection Strategies Super-squad Super-squad 0 0 0 0 0 0 0 0 0 n sources 0 0 0 0 0 hn storage nodes Or collect a large number where n independent equations exist with high probability? Is this matrix invertable? A large collection that guarantees decoding (whp) costs a lot: collection energy constraint TRADE-OFF in collection strategy � Up-front collecting: Collect small super-squad of code symbols locally � fits the energy budget, but insufficient � On-demand collecting : collect selected code symbols � likely to cost more per symbol, but few of them needed during decoding process

Efficient Collection Strategy: Push-Pull Model s { 0 0 0 0 0 0 0 0 n sources n sources { 0 0 0 0 0 Push: The closest super-squad of size s sends coded packets - Enough to decode partially (belief propagation decoder) + − UNSTUCK!!! STUCK! s d n Doping Cost: n Pull: Query for d code symbols which can continue belief-propagation decoding (decoder doping)

Random Fountain Encoding of Network Data Doping Cost Depends on Degree Distribution How to select degree distribution to collect efficiently? desired undecoded rate is 0.01 100 RS constant: 0.10 90 I: Ideal Soliton 80 R: Robust Soliton percentage od doping symbols 70 Robust Soliton 60 50 I min doping percentage I mean doping percentage 40 I max doping percentage R min doping percentage 30 R mean doping percentage R max doping percentage Ideal Soliton 20 10 0 0 500 1000 1500 2000 2500 3000 3500 4000 n: number of symbols to decode

Random-access Push-Pull Data Collection and Decoding Doping Cost Depends on Doping Mechanism How many coded packets do I need to pull to decode all data? LEGEND D: Fountain with Degree-2 Doping U: uniform doping 100 D min DP D mean DP 90 D max DP U min DP U mean DP 80 U max DP 70 percentage od doping symbols 60 50 random packet pull 40 30 20 10 0 0 500 1000 1500 2000 2500 3000 3500 4000 n: number of symbols to decode “smart” packet pull

Thank You.