Tiny Networking

1
Tiny Networking

• David Culler
• University of California, Berkeley
• Intel Research Berkeley

http//webs.cs.berkeley.edu
2
Vast Networks of Tiny Devices

• Past 25 years of internet technology built up
  around powerful dedicated devices that are
  carefully configured and very stable
• local high-power wireless subnets at the edges
• 1-1 communication between named computers
• Here, ...
• every little node is potentially a router
• work together in application specific ways
• collections of data defined by attributes
• connectivity is highly variable
• must self-organize to manage topology, routing,
  etc
• and for power savings, radios may be off 99 of
  the time

3
Directed Diffusion Concept

• Nodes express interest in data with certain
  attributes (sinks)
• Establishes gradient from sources

Estrin, Govindan, Heideman
4
Directed Diffusion Concept

• Nodes express interest in data with certain
  attributes (sinks)
• Establishes gradient from sources
• Sources generate data
• Useful paths reinforced, others suppressed
• in-network aggregation
• nested queries

5
Huge Design Space Application

• Traffic
• any-to-any, all-to-one, one-to-all, collection of
  sub-groups, ...
• steady low-BW steam of readings/findings, bursts
  of data, periodic logs, ...
• Duration
• years (hazard alarm), months (field season), when
  the big red button is pressed
• Available infrastructure
• power, base-stations, location
• Mobility, Stationary
• all fixed, all mobile, mixture, changing
  environment
• Placement / Physical Topology / Scale
• arbitrary, controlled, unknown
• Network Services
• time synchronization, localization, proximity,
  ...
• gt Best way to get a networking solution now is
  to define your application context

6
Design Space Underlying Technology

• Link (Radio) Technology
• range, control of signal strength, noise
  tolerance
• channel capacity, coding, error rates
• single channel, multi-channel, tunable
• Device Density, Failures over time
• MAC (media access control)
• channel sensing, back-off, protocols, collision
  behavior, link-level acks
• time synch, energy aware
• Power management Energy Constraints
• scheduling, functional allocation
• Transmission Rate Control
• Topology Formation
• hierarchical spine, geographic, static, dynamic
• Routing
• single path, multipath, passive participation
• explicit vs implicit nbhd detection

7
Losing the forest for the trees

• Systems side
• clever fix to one particular aspect, but only
  looking at it in sliver of the design space
• Theory side
• assuming cell coverage is a disk of radius r with
  sharp boundary
• within radius Pconnect 1, outside Pconnect 0
• eg., unit disk-graphs for routing, maximal
  independent sets, min. dominating sets, leader
  election
• Wireless communication between small devices is
  inherently noisy, unpredictable,
  non-deterministic
• Radio signal fades with distance
• Interference
• Multipath (reflections)
• Collisions
• Mobility or not

8
Surge Demo
9
Local Operations gt Global Behavior

• Nodes sense network environment
• uncertain, partial information
• Packets directed to a parent neighbor
• all other neighbors hear too
• carry additional organizational information
• Each nodes builds estimate of neighborhood
• adjusted with every packet and with time
• Interactively selects parent
• Routes traffic upward
• Collectively they build and maintain a stable
  spanning tree
• takes energy to maintain structure

node depth child? parent? link goodness
17 1 yes 90 .7
6 3 yes 75 .6
...


Predictable global behavior built from local
operations on uncertain data
10
Wireless Connectivity

• Controlled study on 13x13 array of Rene nodes
• Single transmitter
• Record fraction of packets received at each node
• Many packets at each of several transmit power
  levels
• complex fall-off over substantial range
• range defined by CEP

Contours of probability of reception from center
node for range of transmit power strengths
with Ganesan, Woo, Krishnamacheti, Estrin
11
Alternative Perspective

• Develop algorithms that are built fundamentally
  upon a probabilistic connectivity structure
• Embrace noise, rather than fight it
• the network is really another sensor (and
  actuator)
• Utilize simple, local rules, rather than complex
  protocols
• Challenge obtain predictable global behavior
• Challenge interfaces for imperfect operation

12
Simple Epidemic Broadcast Schema

• Local Rule
• if (new mcast) then
• take local action
• retransmit modified request
• Should forms roughly breadth-first spanning tree
• Examples Network wakeup, command propagation
• Build spanning tree
• record parent
• Naturally adapts to available connectivity
• Minimal state and protocol overhead
• gt surprising complexity in this simple mechanism

13
Network Discovery Radio Cells
14
Network Discovery
15
Behavior at Scale

• Variations in connectivity
• With many nodes, likely that one far away will
  hear
• Long links tend to be asymmetric
• Structure dominated by contention

16
Final Tree
17
Power Laws ?

• Most nodes have very small degree (ave .92)
• Some have degree 15 of the population
• Few large clusters account for most of the edges

18
Open Territory gt Many Children

• Example Level 1

19
Open Territory gt Many Children

• Example Level 2 variation in distance

20
Open Territory gt Many Children

• Example Level 3 long links

21
Importance of Asymmetric Links

• Asymmetric Link
• gt65 successful reception in one direction
• lt25 successful reception in the other direction
• 10-25 of links are asymmetric
• Many long links are asymmetric
• in large field it is likely that someone far away
  can hear you
• what does this mean for protocol design?

22
Collisions are primary factor

• Nodes out of range may have overlapping cells
• hidden terminal effect
• Collisions gt these nodes hear neither parent
• become stragglers
• As the tree propagates
• folds back on itself
• rebounds from the edge
• picking up these stragglers.
• Seen in many experiments
• Mathematically complex because behavior is not
  independent beyond singe cell

23
Probabilistic Connectivity Model

• Radio signal fades with distance in complex
  manner depending on environment
• Radio receiver has complex behavior to extract
  signal
• What matters to algorithms is whether packets are
  delivered or not
• work directly with probabilistic communication
  model
• but which one?
• Calculate comm. rates for numerous
  transmit/receive pairs at range of distances

24
Fall-off of Probability of Comm.
Low Power
High Power
25
Drive Simulation from Empirical Stochastic
behavior
26
Example Cell Coverage from Model
feet
feet
27
Reception Model with Collisions

• Second experiment with two nodes sending at once,
  record which nodes hear which one
• follows P success closely
• When does a second sender collide?
• Clear comm. region gt YES
• silent region gt NO
• transition region gt collides if would have
  communicated
• Reception Model
• Assume pij is the probability of success (i-gtj),
  Probability for B to receive As message
• pab ?icollider(1-pib)

28
Common Special Case Data Gathering

• Collection of nodes take periodic samples
• Stream data towards a root node
• Root announces interest
• depth 0
• Nodes listen to neighbors
• When hear neighbor with smaller depth
• start transmitting data to good neighbor with
  smallest depth
• set own depth to one greater and include with
  data
• Data transmission continuously reinforces /
  adjusts routes

29
Use Case Assumptions

• Application
• N-to-1 all data, no aggregation
• Each node generates small packets at regular
  interval
• Appln phase shift on collision
• Routed to a specific node, e.g, base station or
  root of request
• data rate below saturation
• 7x7 grid, 10 ft spacing gt several hops
• Underlying
• link-level acknowledgement (may be used for
  retransmission)
• CSMA with simple channel sensing, fixed backoff,
  initial random delay
• Receivers always attentive
• may use sampled listening
• Broadcast choose parent first contact
• Shortest Path choose parent closer (which one)
• Min path loss choose parent next step on min
  loss path

30
Max-Path-Reliability routing

• each node maintains estimate of loss rate over
  entire path to root
• select nbr on the minimum loss path as parent
• Pito root through j Plink i,j Pjto root
  
• assuming loss rate along path is independent
  of how packets enter the path
• Subtleties
• estimating link rates
• transient error in link rate may lead to cycles
• rate of updates
• stability vs responsiveness
• congestion
• warm-up phase

31
Broadcast

• overall success rate 18.9

32
SP50 Path reliability

• overall success rate 44.8

33
MPR path reliability

• overall success rate 54.7

34
Routing Distance Distribution
SD50
MPR
35
Distribution of success rates
MPR
SP50
36
Observations

• Better success rate with more, better hops
• Partial information causes temporary loss
• cycles introduced when link error estimates are
  poor
• tracks std. dev. of link error rates
• if difference between candidates not
  statistically significant, choose based on hop
  count
• especially important during warm-up phase
• also after topology changes
• Shortest Path will tend to use most marginal
  neighbors
• although range is highly variable, for particular
  pairs of nodes at particular distances
  connectivity is very bimodal

37
SP75 Path Reliability

• overall success rate 52.7 (vs 54.7 for MPR)

38
Multihop Path-Rate Contours
Min Hop (75 nbr)
Min Path Loss
39
Hop-by-hop retransmission

• Decent neighbors become good neighbors if you are
  willing to chat for a moment
• may choose parent that is on path of least
  transmissions
• expected number of transmissions 1 / P(success)
• Distributed computation is estimate of SUM of
  retransmissions
• A few retransmissions make large difference in
  success rate

40
Overall Results
Success Rate to BS Retrans / App packet Total Msg. Overhead/ pack. Recv. By BS Ave. Hop Count
MRP 83.94 1.03 8.36 6.12
Shortest Path 85.79 0.82 7.13 6.18
Broadcast 42.96 5.76 13.13 4.92
41
Fairness
SP75
MRP
Broadcast
42
Stability

• How often does the routing algorithm change?
• of parent changes per unit time

MRP
SP75
43
Stability (Broadcast)
44
Key Building Block Link estimator

• Nodes assess quality of link from packets they
  hear
• directed to them or snoop
• sequence number gt infer losses between packets
• For each new packet (or empty window) revise
  estimate of link probability
• classic EWMA
• P i1 aP i (1-a) X i , where X i 0,1 if
  loss, success

45
Link Probability Changes Dynamically
46
Revisit Classic Estimator

• Want estimator that is responsive to change, but
  stable with low error
• Candidates
• EWMA stable, agile, flip-flop
• Moving Average
• ...
• Best ended up being the own that we eyeballed
• EWMA cascaded with average over a time window

47
Estimator driven from prob. generator
48
Empirical Trace
49
Bottom Line

• Can only estimate link rate to within 10
• Takes about 100 packets to settle
• Design distributed algorithms with this kind of
  approximation of an inherently noisy world

50
So how about that demo?
51
Higher Level Network Services

• Time Synchronization
• many nodes with drift and offset variation
• classic
• pairwise timesynch limited to half the variance
  in RTT
• NTP hierarchy to establish rough global time
• local timesynch
• deep integration into network stack (Hill)
• multiple receivers, not send-rcv (Ellison, Girod)
• global propagation huge open distributed
  algorithm question
• Ranging Localization
• Naming
• Data distribution, Aggregation, Storage
• Weve just begun to tackle the full problem

52
Conclusion

• The networking part of the future sensor
  networks very different the power-hungry, sparse
  IP world
• severe constraints, noisy, dynamic, partial
  information
• Low power means you hardly ever listen
• self-organized and inherently distributed
• opportunity to work across layers of abstraction
• Presents a rich set of open problems
• Probabilistic connectivity model is usable and
  powerful
• On-line distributed approximations to global
  optimization
• a whole new internet to invent
• Applications define valid assumption sets for
  network design
• guide the front of progress in this huge space

53
Engineering Perspective

• Low-hanging fruit
• enough base stations that only need single hop
• Mid-range fruit
• stationary, well-powered multi-hop backbone
• Research centroid
• unstructured, ad hoc, multihop routing in energy
  starved environment