An Empirical Study of Epidemic Algorithms in Large Scale Multihop Wireless Networks Authored by Deepak Ganesan, Bhaskar Krishnamachari, Alec Woo, David Culler, Deborah Estrin, Stephen Wicker Presented by Tibor Horvath CS851 Fall 2003, University of Virginia 11/17/2003
Outline Introduction Related Work Experiments Analysis of Results Conclusion, Opinions
Introduction A large amount of communication algorithms already exist for Wireless Sensor Networks (WSN-s) Most existing algorithms were validated by idealized simulations or small-scale real experiments This paper describes real large-scale WSN experiments Results show that many previous design assumptions were unrealistic A foundation work to support future algorithm design
Introduction: Contributions Characterizes radio communication properties to be expected in real WSN-s Asymmetric links can be significant at large scale Irregular propagation causes uneven distribution of loss rate over distance Obstacles and collisions can cause unexpected black holes with flooding Our simulators should be capable of modeling these properties
Introduction: Contributions Provides insight into some tradeoffs between different transmission power settings High power can cause very high contention But high power can also save energy by reducing multi-hop communication (network diameter) Low power can increase the number of asymmetric links present But low power also has more regular propagation and much less collisions It would be nice to have measurements of energy usage of nodes with each power setting
Introduction: Contributions Shows experimental evidence of the “broadcast storm” generated by flooding Even at low contention levels, the percentage of useless broadcasts is over 60% Higher radio power makes the overhead even worse: Each node reaches much more other nodes many collisions → long backoffs → late broadcasts Even with this overhead, there will always be nodes that do not receive the flood (stragglers) Do algorithms that rely on flooding really accept or tolerate these properties? How does it affect their scalability and robustness?
Introduction: Contributions Demonstrates the non-optimality of routing protocols that use shortest reverse hop-count paths Asymmetric links alone defeat some of the protocols (e.g. AODV) Long links cause highly clustered trees with low robustness and uneven energy depletion Collisions can cause creation of trees with backward links: routing path flows away from the destination We should be very careful when and how to use hop- counts resulting from flooding in our algorithms
Related Work Prior experimental studies with lack of infrastructure DSR: 8 laptops with , moved in a 300×700m area AODV: 1 desktop and 5 laptops with Data aggregation in Directed Diffusion: 14 PC/104 (Embedded PC) sensor nodes Radiometrix RPC modems: Reliable 30m in-building range, 120m open ground 13.5mm16mm 54mm 32mm
Related Work Prior small-scale experimental studies MAC adaptive rate control: 11 Berkeley motes S-MAC (energy-efficient MAC): 5 Berkeley motes Simulation analyses Realistic modeling is very challenging Cannot be considered as final validation
Related Work Broadcast dissemination algorithms Sophisticated epidemic protocols Probabilistic rebroadcasting (Gossip) Counter-, distance-, location-, cluster-based rebroadcasting Other mechanisms SPIN (energy-efficient) Minimum connected dominating sets (virtual backbone)
Experiment Scenarios: Algorithm Generic epidemic algorithm Retransmission decision is a randomized function of local state Algorithm used for analysis: Flooding Retransmission decision is always true.
Experiment Scenarios: Algorithm Why evaluate simple flooding? Many dissemination schemes still rely on flooding. (e.g. Maté) Although more sophisticated alternatives exist, flooding adequately demonstrates the same physical and link layer issues
Experiment Scenarios: Platform Rene Mote 916 MHz single channel 10 kbps raw bandwidth Dynamically tunable transmission power:
Experiment Scenarios: Platform Rene Mote 916 MHz single channel 10 kbps raw bandwidth Dynamically tunable transmission power:
Experiment Scenarios: Platform Calibration Fresh batteries Same antenna length Vertical orientation TinyOS CSMA with random backoff (6ms-100ms) No packet dropping
Experiment Set 1: Link Characteristics 169 nodes on a 13×13 grid, 2-feet spacing Nodes transmit sequentially to the base station at 16 different power levels Collisions were eliminated About 54,000 messages total Receivers log message data for later reconstruction The results are packet loss statistics
Experiment Set 2: Flood Propagation 156 nodes on a 13×12 grid, 2-feet spacing Open parking lot, no obstacles 8 different power levels Base station in the middle of the grid’s base Nodes log data in all layers for reconstruction of propagation ID of sender → Propagation tree MAC Layer timestamps → Backoff time, Collisions Link Layer timestamps → Minimize receiver delay Reconstruction error under a bit-time per hop
Experiment Set 2: Observations Flood initiated Step 1.
Experiment Set 2: Observations Flood initiated Failed nodes Step 1.
Experiment Set 2: Observations Flood initiated Failed nodes Long links Cell region is far from a simple disc Physical/Link level effect Step 1.
Experiment Set 2: Observations Rebroadcasts Backward links The flood extends towards the source Step 2.
Experiment Set 2: Observations Rebroadcasts Backward links The flood extends towards the source Stragglers MAC-level collisions Step 3.
Experiment Set 2: Observations Final state Backward links The flood extends towards the source Stragglers MAC-level collisions High clustering Most nodes have few descendants A significant few have many children Step 4.
Analysis of Results Physical and Link Layer Effective communication radius Packet loss statistics Bidirectional and asymmetric links Medium Access Layer Contention, collisions Hidden terminal effect Network and Application Layer Propagation structure
Analysis: Physical and Link Layer High transmit power Packet reception map 90% 80% 70% 60% 50%
Analysis: Physical and Link Layer Low transmit power Packet reception map 90% 80% 70% 60% 50%
Analysis: Physical and Link Layer Distribution of packet loss over distance is non- uniform Throughput is lower than 100% even at short distances Insufficient signal processing and error correction Reception decrease not as sharp as signal strength decay (exponential)
Analysis: Physical and Link Layer Distribution of packet loss over distance is non- uniform Throughput is lower than 100% even at short distances Insufficient signal processing and error correction Reception decrease not as sharp as signal strength decay (exponential) Good Bad Neither
Analysis: Physical and Link Layer Connectivity Radius Radius R of the smallest circle that covers 75% of the “good links”: High Med Low V.Low Good link Neither good nor bad link
Analysis: Physical and Link Layer Asymmetric Links: 5-15% “good” link in one direction, “bad” link in the other Bidirectional Links “good” link in both directions
Analysis: Physical and Link Layer High transmit power Very low transmit power Percentage of asymmetric links grows with distance The growth is greater at lower transmit power Small differences in reception sensitivity, hardware, and energy level dominate at the fading edge
Analysis: Physical and Link Layer How would SPEED perform in large-scale? Original experiments use 25 motes on a 5×5 grid Is it sensitive to long links? Can it form backward links? Does it accept asymmetric links?
Analysis: Medium Access Layer Contention Communication range increases with transmit power Interference range is often greater than the communication range Thus, contention increases with transmit power Backoff delay Higher transmit power leads to longer backoff durations However, it is not fully deterministic due to the random backoff implemented in the TinyOS MAC protocol.
Analysis: Medium Access Layer
Maximum backoff interval Captures contention level within interference cells Reflects the largest contention time Approximate, because the starting time of backoffs is not the same among the nodes.
Analysis: Medium Access Layer Reception Latency Definition: The amount of time it takes for each node to receive the flooded packet. Significant fraction of time taken to reach last few (5%) nodes → Stragglers Reception latency increases with network diameter (maximum hop count) → Higher transmit power yields lower latency
Analysis: Medium Access Layer Settling time Definition: Combination of the reception latency and the time taken for all retransmissions to complete throughout the network ReceptionLatency MaxBackoffTime Minimum Settling Time ReceptionLatency MaxBackoffTime Maximum Settling Time
Analysis: Medium Access Layer Low transmit power Settling time is dominated by reception latency because of larger diameter High transmit power Settling time is dominated by maximum backoff time because of high overall contention Nodes keep retransmitting the message long after 95% reached Metric relations: Timings vs. Transmit Power
Analysis: Medium Access Layer Observe the fraction of Reception Latency and Settling Time
Analysis: Medium Access Layer Useless Broadcasts Definition: A rebroadcast that only delivers the message to nodes already reached Note that simple flooding has an implicitly high percentage of useless broadcasts (60%+)
Analysis: Medium Access Layer Collision Appearance of stragglers and backward links can be explained with collisions Stragglers likely form backward links if ever reached later Hidden terminal problem A node is unable to receive most messages due to an obstacle This likely defeats its collision avoidance algorithm Its transmissions likely cause many collisions
Analysis: Medium Access Layer Higher transmit power results in more hidden terminals and thus more collisions
Analysis: Medium Access Layer How much does the high backoff impact the real-time performance of SPEED? High miss ratio What power setting should it choose?
Analysis: Network and Application Layer Dissemination Tree Characteristics Reverse path may fail due to asymmetric links Long links exacerbate this effect as they are more likely asymmetric Long links are likely preferred by applications (routing) Backward links cause suboptimal behavior E.g. sensor data flows away from base station Earliest-first parent selection results in clustered tree Large clusters occur frequently irrespective of transmit power Clustered trees suffer large connectivity loss from orphaning
Analysis: Network and Application Layer Dissemination Tree Characteristics Tree level only loosely corresponds to distance: Stragglers Long links
Analysis: Network and Application Layer How all this affects existing localization schemes? GPS-less... localization paper argues that the idealized radio model “compares quite well to outdoor radio propagation…” Because they use the Radiometrix RPC-s: 120m reliable open ground range in a 10m×10m test area!
Analysis: Network and Application Layer How all this affects existing localization schemes? Range-Free… localization paper assumes an irregular radio pattern But it is still not fully realistic: assumes 100% reception rate within a lower bound distance, 0% beyond an upper bound.
Analysis: Network and Application Layer How all this affects existing localization schemes? Range-Free… Approximate PIT Test: How do long links affect localization accuracy?
Analysis: Network and Application Layer How all this affects existing localization schemes? Hop-count distance based localization schemes The relation of hop-count to distance is very far from being linear:
Conclusion Even simple distributed WSN communication algorithms show very complex behavior Probabilistic connectivity Unexpected links (long, backward) Stragglers Asymmetric links are frequent It is imperative to validate all communication algorithms by performing: Non-idealized simulation based on real data Real large-scale experiments
Opinion: About the results Which properties may be improved? Physical Layer properties? Calibration already near ideal Newer radios (e.g. Mica 40 kbps) have better error correction MAC Layer properties? Rudimentary MAC implementation of TinyOS Directional radios may reduce contention Application Layer properties? Simple Flooding: Broadcast storm problem: Flooding may result in excessive redundancy, contention, and collision. (S.-Y. Ni, Y.-C. Tseng, Y.-S. Chen and J.-P. Sheu, The broadcast storm problem in a mobile ad hoc network) Hierarchical, energy-efficient dissemination protocols
Opinion: About the paper Although results are subject to significant improvements in all layers, the paper: Makes the case for real large-scale validation Shows that only probabilistic communication models will lead to realistic analyses It would be good to explore the transmit power setting tradeoffs further Node energy consumption measurements would have been especially valuable
End Thank You!