140 likes | 272 Views
Message Efficient Termination Detection in Wireless Sensor Networks. Sandip Bapat Anish Arora The Samraksh Company The Ohio State University. Evolution of WSN systems . Scaling of physical sizes and scales in 2004 , ExScal spanned 250,000 sq.m with 1000+ nodes
E N D
Message Efficient Termination Detection in Wireless Sensor Networks Sandip Bapat Anish Arora The Samraksh Company The Ohio State University
Evolution of WSN systems • Scaling of physical sizes and scales • in 2004, ExScal spanned 250,000 sq.m with 1000+ nodes • indoor testbeds and installations of hundreds of nodes (Kansei, Motelab, etc.) • Scaling of software complexity • application requirements outgrow WSN mote resources • sensing, signal processing, routing, localization, reliable transport, network reprogramming, power management, health monitoring, … • sensor networks as reconfigurable fabrics • multiple applications sharing the same physical network • applications spanning multiple fabrics connected via the Internet as envisioned in GENI • Autonomous management support for wireless sensor network applications is critical
Need for Termination Detection • Detecting convergence of WSN protocols • gossip-based protocols like Deluge “run” forever, yet manager needs to know when new program is downloaded on all nodes • Detecting phase termination in multi-phase applications • phase: logical grouping of application tasks executable on a node • in-order requirement: next phase may depend on completion of previous phase • all phases depend on reprogramming • atomicity requirement: all nodes must complete before switching • phases may not be backward compatible • better performance if phase transitions are synchronized
Termination Detection in WSNs • Classical definition: safety and liveness requirements • detection terminated • terminated detection • What’s new?: reactive computation model • energy is critical for low power WSNs • nodes do not compute forever, only in response to external events • WSN application protocol model: • idle active: receive (first) protocol message • active active: receive at least one protocol message every T seconds • active terminated: no messages received for last T seconds • Energy efficiency • low communication overhead • Composability • reusable with different protocols • no dependence on protocol specifics
Existing Solutions • PIF (query) based protocols (SNMS, TinyDB) • build query-collection tree • periodically query entire network for termination status • simple optimization: set-up one time query (trigger) to respond upon termination • Drawbacks • inefficient • require separate spanning structure to be constructed and maintained • require response from every node • prone to message loss • create message burst when application protocol terminates, increasing contention and interference losses
The Reporter Protocol (1/3) BS: is_reporter = 1 1,2: seen_reporter = 1; parent = BS 1: is_reporter = 0 3,4: seen_reporter = 0; parent = 1 4: is_reporter = 1 3,5,6,7: seen_reporter = 1 5,6,7: parent = 4 BS 1 2 3 4 5 • Property 1: Reporter Set is a Dominating Set (DS) over nodes sending application messages • ideal solution: Reporter set is a • Minimum DS • MDS is NP-hard Reporter selection rule: during an application message send, a node becomes a reporter if no reporters have been heard from. 6 7
The Reporter Protocol (2/3) • Report collection • autonomous, efficient structure creation: exploit application traffic • parent = sender of first received application protocol message • similar to Dijkstra-Scholten’s algorithm • Property 2: Constructed structure is a spanning tree • simple, per-hop reliability: acks, buffering & retransmission • Local termination detection • reporters snoop application messages over broadcast channel • quiescent interval T termination in one hop neighborhood • Property 3: Local termination detection by all reporters is sufficient to satisfy safety and liveness requirements
The Reporter Protocol (3/3) • Detecting global termination • Technique 1: use network localization data • compute terminated regions based on known reporter locations • conservative yet safe approach • Technique 2: learn reporter set • nodes notify manager when they elect themselves reporter • manager matches received reports to reporter set • requires twice as many messages • Composability • minimal knowledge of application protocols required • message type: used to overhear relevant application messages • maximum communication delay: used to compute T • Efficiency • maximally exploit application traffic to reduce messaging overhead
Case Study For Evaluation:Network Reprogramming • Reprogramming is a core WSN service • termination detection is critical before switching to new program • 2 popular reprogramming protocols • Deluge: unstructured, flooding based • nodes periodically advertise own program version number • nodes request & receive newer versions from neighbors • randomized transmission to avoid duplicate sends • some nodes acquire new program by simply overhearing • Sprinkler: structured, TDMA based • exploits topology to construct backbone structure • backbone is within O(1) of MCDS • backbone nodes locally compute TDMA schedule for dissemination • although quite different in operation, both protocols satisfy the reactive WSN model
Evaluation (1/3) • Overhead • per sent/received message: 2 extra checks • per protocol invocation: 2 extra assignments • per message: 1 extra bit • ~50 bytes of mote RAM • Performance evaluation • Kansei testbed experiments using XSM motes • 105 nodes in a 15 x 7 grid topology • different network densities simulated by changing transmission power levels • nodes logged all important protocol execution data locally • exfiltrated offline using reliable, ethernet backchannel • compare logged data to Reporter output to measure reliability
Evaluation (2/3) • Efficiency of Reporter selection • Results • Even at lowest density, only 4-7% nodes become reporters • comparable to MDS of these networks • Despite completely different dissemination patterns for Deluge and Sprinkler, Reporter achieves comparable performance
Evaluation (3/3) • Spatial distribution of reporters • Results • selected reporters are uniformly distributed • better load balancing • PIF queries create traffic burst over whole network • existing protocols achieve 50-90% reliability • for reduced (~5%), uniformly distributed load of Reporter, we achieved 98% end-to-end reliability with per-hop acks.
Conclusions • Reporter • provides reliable, efficient termination detection by exploiting reactive WSN model • only requires about 5% of nodes to respond compared to PIF-based approaches • is fully autonomous • creates its own collection structure and detects termination locally • can be composed with a large class of application protocols • does not assume knowledge of application protocol specifics • reprogramming study used same Reporter code for both protocols • only difference was instantiation of message type & communication delay parameters
Thank you! Questions/Comments Email: sandip.bapat@samraksh.com