Energy-Efficient and Low-Latency, Tree Based Algorithms for Convergecast in Wireless Sensor Networks by Sarma Upadhyayula Committee Dr. Sandeep Gupta Dr. Arun Sen Dr. Hasan Davulcu

Outline of Presentation Introduction Problem statement System model Convergecast vs Broadcast Tree based algorithms Data scheduling algorithm Results Conclusion and future work References

Introduction Convergecast – collecting data from all or some nodes to a central node. Flow of data – opposite of broadcast Issues: Energy-Efficiency, Low-Latency, Network Life time, Fault-Tolerance, Scheduling, Data Compression. Wendi[4] – Life time Stephanie[5] (a.k.a Pegasis)– Energy, Life time Bhaskar[6] – Energy (general case) Valli[2] – Energy Chalmerk[7] – Network Density

Problem Statement GIVEN: Wireless Sensor Network (WSN), N nodes – sources, 1 base station – sink. Definitions Total Energy – sum of individual energies. Total Latency –from start to reception ALL packets Life Time – No. of rounds – before k nodes die Energy-Efficient – high energy for communication [4] Low-Latency – time sensitive applications Long network life time OBJECTIVE: Energy-Efficient, Low-Latency and High Network Life Time

System Model (Assumptions) Data Routing nodes – static and synchronized. Graph – positive edges, bidirectional edges, connected. Edges weight – cost (energy) for communication. All nodes can talk to one another – but they don’t. Multi-hop (over single-hop). Data compression factor,  = [0,1], for intermediate nodes Data Communication Channels, (CDMA code, time slot), for communication One transceiver

System Model Energy Model [6] Latency Model [2] i j Energy Model [6] Ei = (Eelec + εamp × rλ) × d Ej = Eelec × d Typical values[6]: Eelec = 50µJ/bit, εamp = 100nJ/m/bit, λ = 2, d = data bits Latency Model [2] Time Slot T = max{d1, d2, …, dn} Network Life Time Model No. of rounds of convergecast before death of k nodes. 1 2 n

Convergecast vs Broadcast Data flow in a tree is bottom up. Packet size varies. Cannot maximize no. of child nodes. Latency is determined by parallel transmissions.* * = the concept of -constraint Data flow is top down. Constant packet size. Tree – maximize no. of child nodes. Latency is determined by longest path.

Convergecast Algorithm (β = 2)

Algorithms’ Performance Energy gain: 5-8% over broadcast; 2% over other Valli03

Algorithms’ Performance (contd…) Latency gain: 30-40% over broadcast; up to 3 times better than Valli03

Tree Based Algorithms Rationale for tree construction Data Compression Factor,  = [0,1] -constraint (ideally, for  = 1, minimum latency for  = e and for  = 0, minimum latency for  = 2) MST & SPT based trees

Latency vs  for  = 0: Minimum at  = 2 Theoretical Dependence of Latency vs  for  = 1: Minimum at  = e (2.71) Latency vs  for  = 0: Minimum at  = 2 Experimental Results of Latency vs  for  = 1: Minimum at  = 2 Latency vs  for  = 0: Minimum at  = 2

MST based trees ( = 2) Consumes minimum energy when data compression,  = 0 5 7 8 4 6 ∞ Network 5 4 MST Energy = 24J Latency = 20 units MST with -constraint Energy = 26J Latency = 9 units 5 4 7

SPT based trees ( = 2) Consumes minimum energy when data compression,  = 1 5 ∞ Network 6 7 4 2 10 7 5 11 SPT Energy = 44J Latency = 16 units 6 10 7 5 11 12 SPT Energy = 45J Latency = 9 units 6

Data Scheduling Algorithm Assumption: Some fixed CMDA codes are given (code, time slot) pair = channel Parent nodes shouldn’t be in the range of another transmitting node Child node shouldn’t transmit when another parent node is receiving Channel i j(≠ i)

Simulation Environment Battery Energy = 10µJ, Data Compression  = [0,1],  = {1,2,3,N} Node Density = 10% to 80% Network Size = 50 – 300 Comparison with Pegasis [5]

( = 1,  = 3, ND = 10%) ( = 1,  = 3, ND = 10%) ( = 1,  = 3, ND = 80%) ( = 1,  = 3, ND = 80%)

( = 0,  = 3, ND = 10%) ( = 0,  = 3, ND = 10%) ( = 0,  = 3, ND = 80%) ( = 0,  = 3, ND = 80%)

( = 1,  = 3, ND = 10%) ( = 1,  = 3, ND = 80%) ( = 0,  = 3, ND = 10%) ( = 0,  = 3, ND = 80%)

Observations and Explanations MST: High  > 0.75: regardless of ND, always low energy Lowest Latency Worst life time. Low  ≤ 0.75: performance drop - energy wise and gets worse Moderate life time. High , low longevity of the packet ??? Concentrated energy expenditure Part of packet exists after 1st hop ??? Due to poor life time of Pegasis

Observations and Explanations PEGASIS: High  > 0.75: Low energy Highest Latency Good life time - drops with increase in network size. Low  ≤ 0.75 steep drop, energy wise Moderate life time but drops rapidly. Greedy construction Chain structure Rotation of chain head Very long routing distance Static nature of MST and SPT

Observations and Explanations SPT: High  > 0.75: More energy than other two but difference reduces with ND High Latency Best life time. Low  ≤ 0.75: Enormous energy savings 4 6 1 network 4 6 1 SPT 11 units 4 1 MST 9 units ??? Accounts for High Latency- High Energy contradiction Shortest path by each packet Considers neighbors energy resource. Therefore energy spreads in the network

Results Summary  Network Density Priority (Latency or Energy) Recommended Algorithm Recommended  >75% Low(<30%) Medium(30-60%) High(>60%) Energy MST or Pegasis  > 3 75% MST for lower range MST if network size > 150 nodes SPT if size < 150 MST if network size > 200 and SPT otherwise. SPT = 2 for MST > 3 for SPT > 3

Results Summary (contd…)  Network Density Priority (Latency or Energy) Recommended Algorithm Recommended  <75% Low (10-30%) Medium (30-60%) High(>60%) Energy SPT  > 3 - Latency MST  = 2 or 3

Contributions and Future Work Our contributions Refute the notion of using broadcast trees Emphasized the importance of  Significance of  in latency Solution provided for a broader problem set network size, node density, data compression, tree structure Issues addressed: Energy-Efficiency, Low-Latency, Network Life Time and their inter-dependency. Future Work Issues NOT addressed: fault tolerance, channel reliability More general case: k sources, 1 sink

References Sarma Upadhyayula, Valliappan Annamalai and Sandeep Gupta, “A Low-Latency and Energy-Efficient Algorithm for Convergecast in Wireless Sensor Networks”, Proceedings of IEEE Global Communications Conference, 2003. Valliappan Annamalai and Sandeep Gutpa, “On Tree-Based Convergecasting inWireless Sensor Networks”, Proceedings of IEEE Communications and Networking Conference, Louisiana, 2003. Imrich Chlamtac and Shay Kutten, “Tree Broadcasting in Multi-Hop Radio Networks”. IEEE Transactions on Computers, Vol. C-36, No. 10, October 1987. Wendi R. Heinzelman, Anantha Chandrakasan and Hari Balakrishnan, “Energy-Efficient Communication Protocols for Wirless Micro Sensor Networks”. Proceedings of Hawaii International Conference on System Science, January 2000. Stephanie Lindsey, Caugili Raghavendra and Krishna M. Sivalingam, “Data Gathering Algorithms in Sensor Networks Using Energy Metrics”, IEEE Transactions on Parallel and Distributed Systems, Vol. 13, No. 9, September 2002. Bhaskar Krishnamachari, Deborah Estrin and Stephen Wicker, “Impact of Data Aggregation in Wireless Sensor Networks”, International Workshop on Distributed Event-Based Systems, Vienna, Austria, July 2002. Chalermek Intanagonowiwat, Deborah Estrin, Ramesh Govindan and John Heidemann, “Impact of Network Density on Data Aggregation in Wireless Sensor Networks”, Proceedings of Internationals Conference on Distributed Computing Systems, Vienna, Austria, July 2002.

 = e (derivation) Number of children =  Data compression =  = 1-α log  N Number of children =  Data compression =  = 1-α Total Latency = ( log N) (d ∑ (α)i) For  = 1, Minimum Latency when  = e For  = 0, Minimum Latency when  = 2 log N i=0 Number of time slots Length of each slot

Notations Notation Definition Initial Value P set of parent nodes root node C nodes with parents in P Φ Adj(p) nodes NOT part of tree and adjacent to p - Cposs neighbors of nodes in P Adj(s) Ci children of node i β β-constraint input (u,v) distance between u and v update(X) update set X after each round reflecting their definitions Rx(p) reception code of p Tx(p) transmission code of p (p) time slot for transmission i set of nodes with reception code i i L leaf nodes in the tree  total number of codes

Convergecast Algorithm Tree Construction (G, s) while P ≠ Φ for all c є Cposs d = ∞ for all p є P if (p,c) < d Λ (|Cp| < β ν |Adj(c)| = 1) d = (p,c) Cparent(c) = Cparent(c) – {c} parent(c) = p Cp = Cp  {c} update(Adj(c)), update(Adj(parent(c))) C = C  {c} P = C, C = Φ, update(Cposs)