260 likes | 415 Views
Attivita’ svolte sul progetto PATTERN. Rome University ‘La Sapienza’ Chiara Petrioli Firenze, 12 ottobre 2004. Flat vs. hierarchical. Ns2 based comparative performance evaluation of: Directed Diffusion Directed Diffusion Greedy DCA + Directed Diffusion Greedy
E N D
Attivita’ svolte sul progetto PATTERN Rome University ‘La Sapienza’ Chiara Petrioli Firenze, 12 ottobre 2004
Flat vs. hierarchical • Ns2 based comparative performance evaluation of: • Directed Diffusion • Directed Diffusion Greedy • DCA + Directed Diffusion Greedy • An improved flat approach we introduced (Directed Diffusion Light) • Pro and cons of a hierarchical v. flat approach wrt • Possible reduction of the Directed Diffusion overhead when only applied over the backbone • Effect of different kind of aggregation
Directed Diffusion and Directed Diffusion Greedyc S2 S2 Sink Sink S1 A S1 A Interest propagation Gradient formation DD DDG S2 S2 Sink Sink S1 A S1 A EXPloratory_DATA Reinforced gradients
Clusterheads 1 Gateways Backbone Ordinary nodes 3 9 5 6 2 8 1 Distributed Clustering Algorithm • + DCA: each node has a weight (e.g. residual energy). Nodes with the highest weight in theri neighborhood are clusterheads. A node waits to be contacted by neighbor clusterheads with bigger weight, if such clusterheads do not exist it becomes clusterhead itself. • +DD only over the Backbone • DD overhead reduction • +Clusterheads are natural aggregation points
Geografic aggregation RA Total aggregation Aggregation techniques +Data Traffic (thus energy) reduction - Latency increase 1 2 A 1 2 3 3 Point of aggregation
Simulation Scenarios and Results • N=200 sensor nodes with transmission radius equal to 30m and sensing range equal to 10m are deployed in a squared area of side L=200m; • Nodes monitor targets (5), moving in the area according to random waypoint, pedestrian speed. IST EYES prototypes energy model. • Timeout for aggregation: 2s; two events must occur at most within an area covered by a circle of radius 5m to be geographically aggregated. • Metrics: energy consumption, latency, overhead, throughput,… • Results: DDG has a high overhead (70%) • A hierarchical approach able to reduce to 1/3 the energy consumed over the simulation time, and significantly reducing the probability of collision • Improvement not only due to the reduction in overhead obtained by running DDG only over the backbone BUT also to the aggregation • If properly designed both flat and hierarchical approaches can be effective for aggregation
Improving over Directed Diffusion • Many routing protocols rely on flooding of information • Sensor Networks: Directed Diffusion (DD) • Communication in Wireless Sensor Networks (WSN) is inherently one-to-all or many-to-one (independently of the particular routing protocol adopted) Upon detecting events matching a given Interests sources send packets back to the sink The sink propagates interests ‘send me information on pink elephants every 10min’ Idea to improve DD: more efficient one-to-all/many to all data dissemination
Our targets • Low overhead • Need of a small amount of information (one-hop neighborhood knowledge) for the data dissemination process to operate, keep to the minimum the amount of extra information transmitted. • Min. percentage of nodes involved in the transmission/reception of the message (reception of the message might happen several times). This affects • Network load, collision probability • Energy consumption • Reliability (high percentage of reached nodes) • Scalable, simple, resource saving
A geometric random graph model… • Given are n points distributed uniformly at random in a unit square • Nodes have a fixed transmission range equal to r, there is an edge between two nodes iff they are in visibility (i.e. their euclidean distance is <=r) • c is a parameter named local connectivity • We select a subgraph Gc of the topology graph Gr as follows: • Each node randomly selects c among its neighbors • A link (u,v) is included in Gc if at least one among u,v selected the other • What is the likelihood that Gc is connected and how are Gc connectivity properties related to those of Gr? • Extensive simulations show that for c>=4,5 Gc has the same global connectivity properties of Gr. • Number of connected components • Relative size of the giant component Are the same for varying nodes densities, and simulation scenarios BUTGc has a much more reduced number of links over Gr reduced load, reduced energy consumption
Directed Diffusion Light • Rationale • Idea 1: local rules for generating a sparse virtual topology (Gc) • Perform Directed Diffusion over this virtual topology • How to implement it? • Comes almost for free by exploiting hello messages…
Simulation Scenarios • N<=300 nodes with transmission radius equal to 30m are deployed in a squared area of side L=200m; • Two nodes are neighbors in the visibility graph iff their Euclidean distance is less or equal to 30m (unit disc graph assumption). • Nodes deployments: • Connectivity properties of Gc and of the original topology are compared • Interest and exploratory data dissemination overheads are compared. Hill Uniform
Summary of the Results • For the visibility graph Gr and for Gc, c>=4 global connectivity properties are basically identical, independently of the density. • However the number of traversed links is significantly reduced Number of links Number of nodes (<=300), c=1…4
Relative size of the Giant Component (Un. deployment) Relative Size of the giant component Number of nodes, c=1,…,4 • For the visibility graph Gr and for Gc, c>=4 the plots are basically identical, independently of the density.
Number of connected components (Unif. deployment) Number of connected components Number of nodes, c=1,…,4 • Gr and Gc have the same behavior as far as global connectivity is concerned it does not pay off to set up all possible links. The virtual sparse topology generated by our local rule is enough to maintain global connectivity.
Reliability vs. c Percentage of unsuccessfully received packets (worst case, most far away node) 0.25 0.05 Number of nodes, c=1,…,4 • Connectivity parameter c tuned to 4
Summary of the Results • Ns2-based performance evaluation show that DDGL decreases the energy consumption (1/2 what consumed in case of DDG) • Is able to perform effectively aggregation • Significantly reduces the overhead 4-5 times • Without affecting reliability
Topology Control • Two major ways to TC: • Controlling the power at the node to create energy effective topologies • Taking advantage of the network density for turning off the radio interface • “Node sleep” saves a lot • Only a (connected) fraction of the nodes stays up for performing network functions
Geographical Adaptive Fidelity • GAF: Typical example of topology control • Nodes are placed in grids of side r (needs location awareness) • The side r is a function of Tx radius R: If r ≤ R/√5 then every node has a neighbor in an adjacent grid • Idea:Routing-wise all nodes in a grid are equivalent (not really true…) • A grid leader is elected and stays on • Based on the node residual energy • All other nodes in the grid go to sleep • Leaders form a (connected) backbone • Load balancing is realized by periodic re-election of the leader
If node 4 becomes a leader in A, then X and Y are no longer connected GAF, Drawbacks • Small grids (given the relation between r and R) • Even at high density, most grids have only one node (always > 50%) • Results in disconnected topologies: Many grids are adjacent to empty grids (> 80% in sparse networks)
Topology Control for WSN • Selection of sensors for building a connected backbone • Backbone sensors form a dominating set of the sensor nodes (every backbone node is neighbor or a non backbone node) • Inspired by the Distributed Mobility-Adaptive Clustering (DMAC) algorithm • Pros: • Distributed clustering set up and maintenance • Takes care of the mobility of the nodes • Selects backbone nodes based on their “fitness” to perform • Needs no synchronization • Cons: • Designed for general ad hoc networks • Thought of for mobility rather than “static dynamicity” • No awake/asleep mode, no load balancing • New nodes can trigger a “chain reaction” • Heavy on transmissions (“hello,” etc.)
Sensor-DMAC • DMAC with a sensor twist • Key idea: Reduce the Tx for set up and maintenance to a minimum • Limited maintenance, parametrically defined (only when needed, due e.g. to nodes depleting their energy, limited use of hello messages, promiscuous mode used for sake of discoveryng the nodes when entering the network limiting network load) • DMAC weight selection of backbone nodes based on residual energy • Only backbone nodes are awake, all other are asleep
Simulation Results • Metrics (all averages): • Size of generated backbones • GAF disconnections • Number of control bytes sent (per node) • Power consumption (per node) • Parameters of ns2-based simulations • Nodes: ≤ 300 TR 1000 RF Monolithics • Tx range: 30m • Sensing range: 10m • Initial energy: 50J • Tx, Rx, idle, sleep power: 14.88, 12.5, 12.36, 0.016 (mW) • Area: 200 x 200 m • Sensor tx data rate: 1pkt/20s • Packet data size: 64byte • Targets: 10, random waypoint ((0,1] m/s, 20s pause)
Backbone Size • S-DMAC overcomes GAF limits • High percentage of backbone nodes (awake) • Possible network disconnections • S-DMAC produces smaller backbone than GAF • In dense networks (300 nodes) S-DMAC backbone is 44% smaller than GAF backbone • Implies improvements in energy consumption • S-DMAC has NO disconnections while it often happens in GAF
Control Message Overhead • Comparison with DMAC • Accounts for cost of backbone re-organization • Control message reduction ranges from 86.13% (100 nodes) to 94.13% (300 nodes)
Power Consumption • Decrease in power consumption of S-DMAC with respect to DMAC • DMAC uses a lot of “hello” messages and has no awake/asleep mechanism • Improvements (per node) vary from 47.5% in sparser topologies (100 nodes) to 63.67% in denser ones (300 nodes)
Future activities • Working with Universita’ di Ferrara on GeRaF • Thorough performance evaluation • Identification of improved versions • Implementation (?) of a small scale test-bed integrating interest dissemination and simple flat sensor-sink data dissemination schemes • Refinements of localized techniques for generating sparse virtual topology/topology control