communication and coordination in ... -...
Post on 04-Jul-2020
0 Views
Preview:
TRANSCRIPT
11/2/2007
1
Communication and Coordination in Wireless and Sensor Actor Networks
Melodia, Pompili, Gungor and AkyildizPresented By: Adam Browning
6 NOV 2007
Outline
• Motivation
• Related Work
• DEPR
• Actor/Actor Coordination
• Conclusions
• Future Work
11/2/2007
2
Motivation
• As distributed control systems, WSANs need fast, reliable communications
• Permit real-time coordination and cooperation in WSANs
• Maximize energy efficiency and reliability in WSANs
• Define an event-driven partitioning scheme to achieve goals
Related Work
• I.F. Akyildiz and I.H. Kasimoglu [1] defines research challenges in WSANs
• R. Vedantham, Z. Zhuang, and R. Sivakumar [7] defines hazards and avoidance mechanisms in WSANs
• SPEED [8] provides real-time communications• MMSPEED [9] extends SPEED with the ability to
meet different delay & reliability requirements• [10-12] work to guarantee scalability & efficiency
with topology-dependent partitioning
11/2/2007
3
Definitions
• WSAN: Wireless network of sensors and actors
– Actors are not strictly actuators; they also send/receive network messages
• Sender/Actor Coordination: establishing data path between sensors & actors
• Actor/Actor Coordination: actors working together to decide how to handle an event
Definitions continued…
• B: Latency Bound
• Expired: Any packet that is not received within B (also called unreliable)
• r: Reliability ratio
• rth: Minimum acceptable reliability ratio
• Sv: The set of all nodes (sensors & actors)
• SA + SS = Sv; P = {(s,a) : s Є S, a Є A} (set of source-destination connections)
11/2/2007
4
Equations
Constant energy cost / unit distance
Path loss exponent (2 < α < 5)
d Constant energy cost / unit distance
1 iff link ij is associated with k
Cost of ij (E for ij)
Penalty for unreliable packetsQ Number of unreliable packets
Distributed Event-driven Partitioning and Routing (DEPR)
• Relies on local information
• Makes greedy routing decisions
• Group feedback to avoid overhead of providing reliability feedback to sources individually
• State-machine based
• Forms a data-aggregation tree (da-tree)
• Dynamically modifies transmit power of nodes to maintain reliability threshold
11/2/2007
5
DEPR Assumptions
• Each sensor knows where it is
• Each sensor knows the location of its neighbors and the actors
• Network is synchronized
• Nodes can dynamically alter their transmission power
DEPR FSM
11/2/2007
6
DEPR Start-Up
• Entered when packet received or event observed
• Selects next hop based on two-hop rule
– Find Minimum
– Eelec shows up 4x because we send from 2 nodes and receive from 2 nodes
– Guarantees no loops
DEPR Speed-Up
• Tries to improve reliability by minimizing # hops from source to actor with Greedy Routing Scheme
• Sets nextHop to be the node with the greatest absolute positive advance toward an actor
– Algorithm does so by selecting the node farthestfrom current node that is closer to the actor, but within transmission range
– Does the algorithm always work?
11/2/2007
7
DEPR Speed-Up Mess-Up?What about:
A SD1D2
d = 5
d = 4
d = 3
DSR Speed-Up Mess Up
A SD1
D2
?
11/2/2007
8
DSR Speed-Up Mess Up
A SD1
D2
X
DEPR Aggregation
• Basically the opposite of Speed-Up
• If reliability goes above threshold, then find a closer neighbor to save energy
– Leaf nodes start sending to nearest neighbor that is closer to its actor than they are
– Non-leaf nodes start sending to the nearest neighbor on the same tree that is closer to its actor than they are (otherwise could form cycles)
11/2/2007
9
DEPR State Transitions
• Driven by actor feedback
• Actors periodically calculate:
– Reliability at instant k:
– Short-term reliability at instant k:
– Anticipated reliability:
• Does long and short-term reliability calculations to spot trends
DEPR State Transitions cont.
• If predicted reliability is below threshold, tell nodes to move to Speed-Up state
• If predicted reliability is above threshold– And long and short-term reliability are above
threshold tell nodes to go to Aggregate state
– Otherwise keep status quo
• If reliability is below threshold, short-term reliability is below high-mark and predicted reliability is above low-mark, tell everyone to move to Speed-Up
11/2/2007
10
DEPR State Transitions cont.
• If short term reliability is below low-threshold, tell everyone to speed up
– This takes precedence over ALL of the other rules
• If none of the preceding requirements are met
– If expected reliability above threshold, tell nodes to move to Aggregate state
– If expected reliability below threshold, tell nodes to move to Speed-Up state
Recovery State
• Recovery state is entered when a node cannot get a message closer to its actor
• Nodes transmit at highest power, with a virtual proximity so that their neighbors can hear
• If a node hears a neighbor with a higher virtual proximity transmitting to the same destination, then it adds that neighbor to its detour path
11/2/2007
11
Actor/Actor Coordination
• Actors have an area of effect
– Areas of actors may overlap
• Sometimes an actor can’t act on an event
– e.g. camera can’t turn to record the event
• δ[s] , the action completion bound, is the max length of time from when an event is sensed to when it is acted upon
Actor Area Overlap
• Areas 1-8 are overlapping
11/2/2007
12
Actor/Actor Coordination Definitions
• SA = Set of actors
• SC = Set of collectors
• = h non-overlapping areas for which actor c is responsible
• = m overlapping areas for which actor c is responsible
• Hc = Number of non-overlapping areas
• Mc = Number of overlapping areas
Overlap display
• Yellow areas are
• Red areas are
11/2/2007
13
More Definitions
• is the set of actors that can act on the mthoverlapping area for which c is responsible
• Ra[m] is the action range of actor a
• is the maximum power that a can use to perform an action
– Selected from L discrete power levels
– Using more power gets the action done more quickly
• = Efficiency of actor a
• is the time for actor a, workingindependently to complete an action on the mth overlapping area using power level p
• is the power-time relationship for actor a
• K is a constant (in Joules/m2); is the pth
power level for the mth overlapping area
11/2/2007
14
ha = 1 iff actor a is involved in an action; = 1 iff m in a’s action range
Goal is to Maximize: Within Constraints
Energy must remain non-negative
Energy use comes from all areas on which an actor works
Only 1 power level/actor & area; Every area must be acted upon by someone
Multiple actors working together get the job done faster
ha is 1 if a is acting on an area
An actor can only act on events in its action area
Actor/Actor Coordination using Localized Auction Protocol
• Three possible roles– Seller: collector for an event area
– Auctioneer: conducts the “auction” for an overlapping area; selected by the Seller
– Buyer: Actors that can act on an overlapping area
• Buyers make a bid consisting of their power level, time to complete action and remaining power
• Multiple Auctions occur in parallel
11/2/2007
15
Localized Auction Protocol cont.
• Seller selects an auctioneer for each overlapping area in its area of effect
– Seller sends the auctioneer the time that it can be spent acting on the event and the max time to spend on the auction
• If there’s no overlap, then the actor for the area does the work
CONCLUSIONS!
• Sensor/Actor Optimization problem modeled in A Mathematical Programming Language (AMPL) and solved with CPLEX
• DEPR Convergence analysis simulated
– Assumed ideal feedback to focus on DEPR & avoid noise from communication mechanism
– LOTS of details, not critical to results
11/2/2007
16
• Optimum cost is negligible
• Aggregation from speed-up is a growing curve– Aggregation flattening the speed-up cost
• Start-Up grows quickly– Have to send messages further
• Speed-up spikes up– Trading energy for speed
• Why is aggregation from startup is a flat curve?– Data “fusion” consolidates messages to save
energy
Sensor/Actor Energy Cost as a Measure of Event Ranges
11/2/2007
17
Startup Cost as a function of event distance and # of nodes
• Longer distance causes greater cost
– Obvious reason, having to send messages further costs more energy
• More nodes cause greater cost
– Total cost, more sensors bring more messages
• The Speed-Up state tells the same story as Startup, and for the same reasons
11/2/2007
18
DEPR Convergence Analysis
• If reliability is really low, set chances of moving to speed-up to 50%
• If reliability is a little below the low threshold, use a 10% chance to moving to speed-up
• If probability of moving to aggregate state is over 10%, instabilities result and observed reliability plummets
DEPR Observed Reliability
11/2/2007
19
DEPR Convergence
Distribution of Delays
11/2/2007
20
Delays at simulation times
Actor/Actor Coordination
• Tested with three solution approaches
– Optimal: Actors chosen by solving Residual Energy Maximization problem
– One-Actor: Action is taken in overlapping areas by the one actor who will have highest remaining residual energy
– Localized Auction: Approach described earlier
11/2/2007
21
Actor Types
• Tested with two types of actors
– Homogeneous Actors: All actors have efficiency of 0.8
– Heterogeneous Actors: 50% of actors have 0.6 efficiency, rest have 0.9 efficiency
Average Residual Energy (Homogeneous Actors)
11/2/2007
22
Average Residual Energy(Heterogeneous Actors)
Future Work
• Create an accurate, non-restrictive analytical model for the delay of large-scan WSANs under different conditions
• Run simulations longer to see how performance holds up as system runs out of energy
top related