1
Fault-Tolerant Data Collection inFault-Tolerant Data Collection in Heterogeneous Intelligent Monitoring Heterogeneous Intelligent Monitoring NetworksNetworks
Jing DengDepartment of Computer Science
University of North Carolina at [email protected]
http://www.uncg.edu/~j_deng/
Joint work with Profs. Meikang Qiu and Gang Wu
2
Wireless Networks
• Networks formed by wireless devices– All communications are sent through wireless channels.– Wireless devices with limited resource
Battery energy, memory space, computation power
• Many interesting problems:– How to lower communication/computation cost for network
activities?Communication takes time/energy.Computation requires memory space and energy.
– How to protect systems from node failure?Small wireless devices could easily fail or run out of battery.
3
Failure Models
• Fail-stop– Device simply stops working.– No information will be sent or received.– Similarly to one with dead-battery
• Byzantine failure– Device can virtually do anything that it is capable of
• Dropping packets from others• Sending out fabricate packets• Modifying packets from other nodes• Deviating from communication protocols
– Much more difficult to address
We will use the fail-stop model
4
Intelligent Monitoring Networks (IMNs)
• Wireless sensor networks– Networks with (possibly numerous) wireless micro-sensors
• A special type of wireless sensor networks– Likely to be deployed for building structure monitoring, forest
monitoring, levee monitoring, industrial plant monitoring, etc.
• Two key characteristics– Node failures expected– Heterogeneous architecture
• Mostly small devices to collect/report data• Some larger and more powerful devices to process/fusion data• These power nodes send results to observer (data sink).
6
Low-power Low-Cost Devices
• Devices usually use low-power transceivers– Goal: to lower energy consumption and to extend lifetime
• Forming multi-hop communication topology– Relying on other devices’ help to deliver data
• Interference can easily disrupt communication– Network topology changes– Data collection paths change– Data loss
7
BitTorrrent - P2P File Sharing Technique
• Swarm: collection ofnodes with the file(even partially).
• Tracker providesswarm information
• Client downloadspieces from nodes inswarm.
• At the same time,uploading pieces toother nodes
• Finished clients serveas seeds (upload only)
Even when some of the nodes inthe swarm fail (or left), filesharing continues.
8
BitTorrent Strategies
• Two strategies in BitTorrent make it surprisingly efficient– Optimistic un-choking– Rarest-first
• Optimistic un-choking is the strategy to choose peers todownload pieces– Suppose there are 100 peers (with ever-changing D/L speed).– Which of these 100 peers should the client choose?
• Using all of them is impractical.• Choosing the top N peers w.r.t. download speed (N=5)• However, there might be new peers offering higher speed.• -> dropping one of the current N peers and randomly testing
another peer (un-choke one of the unselected peers)
– Benefits• Utilizing most of the peers with highest D/L speeds.
9
BitTorrent Strategies (Cont’d)
• Rarest-first strategy governs how to choose pieces fordownload– Suppose a peer has M of the pieces– Which of these pieces should the client choose?
• Random selection or sequential selections?• -> Always choose the rarest piece among all peers (requiring piece
information from other peers).• So that this piece can be offered to other peers.
– Benefits• Increases piece redundancy• Maintaining torrent health• Improves chance of successful download
10
IMN and BitTorrent?
• Data collection in IMNs shares striking similarities withP2P file sharing
Peers may go offline withoutwarningNodes may fail at any time.
Redundancy of file pieces amongpeers
Monitoring data redundancyamong different nodes
A client tries to download a fullset of pieces from a swarm of
nodes
Observer (data sink) tries tocollect data from monitoring
nodes, which generate the data
P2P File SharingIMN
11
IMN - Connectivity Overview
• Lines connect nodeswho can hear eachother (N=100).
• Darker squares markmore powerful nodes(M=10).
• Result of randomnode placement
12
Fault Tolerant Data Collection
• Powerful nodes collect data from regular nodes– Announcements are made from the powerful nodes.– Multiple trees are formed with data forwarding nodes.
• Usually data forwarding nodes only need to forward datafrom nodes on their own tree– In order to provide fault tolerance, they will choose α of other
overheard transmissions– α is termed support ratio
13
Illustration of Data Collection
• Multi-level datacollection
• We show the [avg,max] record on thepowerful nodes
• α=0.4
• Some nodes fail(marked with red x)
• Big red dot representsa fire burning
• None of the powerfulnodes sees anytemperature anomaly.
14
Illustration of Data Collection (Cont’d)
• The same topologyand data collectiontrees.
• α=0.4
• With the same failednodes, two powerfulnodes receive thetemperature anomaly(Nmax=2).
15
Performance Results - Reading Abnormal Temp.
• Similar simulationswere run and averageNmax computed
• pe is node failureprobability
• Nmax lowers as peincreases.
• With larger α, Nmaxincreases.
16
Data Loss due to Failed Sensors
• Failed nodes lead todata loss
• Support ratio α candramatically reducedata loss.
17
Conclusions
• Wireless monitoring networks can provide robustenvironment monitoring.
• We have proposed a fault-tolerance data collectiontechnique for IMNs:– Multiple multi-level data collection trees (forest)– Data forwarding nodes process overheard data.– Support ratio α
• Benefits of our scheme have been demonstrated– Low cost– Fault tolerant toward node failures