DESYNC: Self-Organizing Desynchronization and TDMA on Wireless Sensor Networks (IPSN’07) Julius Degesys, Ian Rose, Ankit Patel, Radhika Nagpal Division of Engineering And Applied Sciences, Harvard University

Introduction Synchronization is biologically–inspired: cardiac cells beating together, fireflies It is an important primitive in WSN: decentralized algorithms for node synchr so that the network remains stable despite individual faults / topology changes

Desynchronization Is the logical opposite of synchronization Nodes performs their tasks as far away as possible from all other nodes The goal is to interleave periodic events to occur in a round-robin schedule: imagine not fireflies flashing in unison but in a uniformly distributed fashion

Desynchronization Setting: Motes performing periodic tasks Synchronization: Sometimes desirable to perform tasks at the same time

Desynchronization Setting: Motes performing periodic tasks Synchronization: Sometimes desirable to perform tasks at the same time Perfect Desynchronization: Other times, want even distribution No two are happening at the same time Perfect round-robin schedule Wireless spectrum not “2.4GHz”

DESYNCH A biologically inspired, self-mantaining algorithm to achieve desynchronization in a single-hop network. Possible applications: equally distribute monitoring tasks organize sleep cycles improve MAC protocols such as TDMA

Background - Synch modelling Many synchronized systems are modeled as networks of oscillators : each node pulses at a fixed frequency existence of an algorithm converging to synchrony (Mirollo-Strogatz, Synchronization of pulse-coupled biological oscillators, SIAM 1990) Desynch can be considered as a patterned synch where all oscillators pulse at evenly spaced intervals: algorithm converging to desynchrony (DESYNC) in O( n^2), for n < 500

Background 2 – Channel sharing in WSN In WSN nodes needs to share the channel MAC protocols: Contention-based (CSMA/CA): nodes checks the channel before transmitting if busy they randomly back off and try later Schedule-based (TDMA): time is divided into slots, each node picks one of them to transmit collision-free

CSMA/CA Pro: simple and adaptive Draws: used when the expected contention is low / in case of bursty traffic Draws: unable to detect collision while transmitting hidden terminal problem large backoffs and message loss under high load low bandwith utilization

TDMA Pro: Two main drawbacks: low latency & good bandwith (fixed frequency) good for predictable traffic (fixed bit rate) Two main drawbacks: overhead: tyme synch among nodes & slot schedule negotiation wasted slots: slot goes unused when node does not need to transmit

Key Observation TDMA only requires nodes to desynchronize their transmissions There is no need to agree on global time If nodes could self mantain desynchronization, we can solve the above drawbacks simultaneously DESYNC applied to TDMA

The Framework Single-hop network Each node performs a task with period T Фi(t)= phase of node i at time t, 0 ≤ i ≤ n-1 Nodes moving clockwise on a ring Upon reaching Фi(t)= 1 node i “fires”, indicating the end of its cycle to others

All nodes observe the firing, and: use this information to jump forward or backward in phase Being Δi(t)= Фi(t) – Фi-1(t): GOAL: Δi(t)= 1/n , 0 ≤ i ≤ n-1 Desync Nodes oblivious of the current state, aware only of the firing events

Framework ? How do we get from a random start to desynchronization? Single node periodically “fires” by broadcasting a message A typical starting configuration (random) A system in desynchronization How do we get from a random start to desynchronization? ? Don’t use the word “Sample,” broadcast a message to let other nodes know that it is performing its task This represents the progression of the node’s internal timer.

DESYNC Algorithm Local view of three nodes, let’s focus on B FIRE A A B B C Local view of three nodes, let’s focus on B Node C has already fired, as well as B. B has tracked C’s firing time. Node A is about to fire. When back neighbor fires, B jumps towards the midpoint of the nodes preceiding and following it A node pay attention to the timing of the firings before and after its own (phase neighbors)

DESYNC Algorithm(2) Δi+1, Δi = firing time of previous and next nodes relative of node i’s firing Node i can compute the phase of its previous neighbor: Фi+1(t)= Фi(t) + Δi+1 (mod 1) and the one of its next neighbor Фi-1(t)= Фi(t) - Δi (mod 1) note Δi+1 it’s in node’s memory (stale)!

Using that info, node i computes the desired midpoint as: Фmid(t)= 1/2 [Фi+1(t) + Фi-1(t)] Now he can jump towards it: Ф’(t)= (1-α) Фi(t) + α Фmid(t) where α Є (0,1). It turned out that α = 1 does not always allow convergence ( oscillation due to memory time delay)

Simple Implementation: constant memory regardless of the network size DESYNC Algorithm(3) Convergence to Desynchrony: regardless the usage of stale informations. Simple Implementation: constant memory regardless of the network size Self-Adapting: if number of nodes changes, those closest to the disturbance adjust their phases, leading the system back to desynchronization.

DESYNC-TDMA Algorithm Nodes use earlier firings to compute the TDMA slots near the time of their next firing Node i’s TDMA slot begins at previously computed midpoint between i and its previous phase neighbor It ends at the midpoint between i and its next phase neighbor

A node will never fire outside its slot: If node B doesn’t jump: ФA < ФB < ФC ФA+ФB < 2ФB < ФC+ФB mid(AB) < ФB < mid(BC) Being B’ = mid(AC) the target jump point of B, this point is always within B’s time slot: ФB < ФC ФA < ФB ФA+ФB < ФA+ФC ФA+ФC < ФB+ФC ФA+ФB < ФA+ФC < ФB+ФC mid(AB) < mid(AC) < mid(BC)

DESYNC-TDMA Key Features 1)The algorithm fully utilizes the channel defining a set of non-overlapping slots covering T Collision free transmission, fully utilized bandwidth

DESYNC-TDMA Key Features(2) 2) TDMA schedule adapts to nodes entering or leaving. When a node leaves, its neighbors readjust their slot boundaries if node does not need to transmit: leave the protocol, sleep & re-enter When a node enters he must interrupt: cost of one time period latency / one data slot interrupted

DESYNC-TDMA Key Features(3) 3) The algorithm is self-contained: No need to know the network size No need to discover neightbors IDs No need to agree on a global time No need for a time synchronization protocol The round-robin schedule is a result of the order in which nodes enter the process

DESYNC-TDMA – Implementation Implemented on Telos WSN with TinyOS Motes use a 250Kbps wireless transceiver. Default CSMA radio interface to transmit, with initial backoff reduced to 1.2 ms Motes use their local clock to track firings MAC-level TS used to insert a delay into msg

Evaluation – Experimental Setup Single-hop network, 20 Telos motes around a Tmote Sky acting as a base station The BS logs all messages

Two experiments: Fixed-Size: a fraction of total motes (n = 4,10, 20) to transmit data in the entire slot, to test TDMA-like performance under peak load. Node Removal-Addiction: in order to evaluate the effect of motes entering- leaving. Initially n = 8, then a mote removed at t = 135. At t = 180 three motes enter the system.

Evaluation Metrics Average Desync error: is the average deviation from desired slot-size of T/n for a given round Normalized throughput: best possible data throughput is 62.8 Kbps (single mote to BS). Ratio between measured data msg throughput and the above value, meaured at each round Fairness: average, min & max throughput per node Message Loss: BS detect missed msg by SN Ratio between missed msg and expected msg.

Experimental Results Motes arrival and departure experiment: 8 motes started At t= 135 one mote left and three motes woke up at t=180 Time of each mote’s firing relative to those of a single mote ( fixed-size experiment, n= 10) Desynchronization error over time for different n Desync error and throughput changes over time ( fixed-size experiment, n= 10)

Comparison to other MAC Protocols Ideal TDMA: upper bound, collision-free slots of T/n size, fully utilized bandwidth. Fixed TDMA: lower bound, slot size of T/N Hybrid TDMA: (Z-MAC). TDMA is modified so that unowned slots can be used by other motes. CSMA: simple and adaptive, works well for small networks and variable traffic. Large backoffsand message loss under high load.

Comparison to other MAC protocols Throughput and message loss across different protocols Normalized throughput over n, for different protocols where each node tries to send as much as possible (high data rate) Total throughput on a 10 nodes network varying data rates

Summary and Conclusion Desync-TDMA is a new way of thinking about TDMA scheduling. Draws: - 1 round latency before transmitting - Smaller slot for several round until re- achieving desynchronization - Fairness could lead to wasted bandwidth - Effect of lost firing messages on desynchronization accuracy? Pro: - No explicit scheduling / time synchronization - Excellent total throughput and collision- free transmission under high loads - Fairness and predictable message latencies - Self-adaptive to topology changes

Future Work and Open Issues Applying DESYNC to coordinate sleep schedules Extend DESYNC-TDMA to multi-hop networks. This is much more complex due to: Intersecting neighborhoods Overlapping broadcast regions: hidden terminal

Any question?

Thanks for your attention