Trickle: Code Propagation and Maintenance Philip Levis UC Berkeley Neil Patel UC Berkeley David Culler UC Berkeley Scott Shenker UC Berkeley ICSI
Retasking a Sensor Net Long lifetimes require retasking Spectrum of retasking mechanisms Binary images (MOAP, XnP, Deluge) High-level virtual programs (Maté, TinyDB) Parameter setting (Attribute, Global) Install on every node Packet loss and transient disconnection Periodically check that everyone has the right code (advertise) TinyDB tuples embed query ID NEST Retreat, Jan 2004
Propagating Can Be Costly, Binaries: 10-60KB Virtual programs: 20-400B Parameters: 8-30B To every node in a large, multihop network… NEST Retreat, Jan 2004
Knowing When Is Costlier Periodically checking that everyone has the right code (advertisements) can cost more than the code itself. 64KB of data: ~1 packet/minute for a day 400B of data: ~1 packet/hour for a day 20B: one packet! NEST Retreat, Jan 2004
Problem Statement The first step is to detect when nodes need updates (continuous process) When there is no new code Maintenance cost should approach zero When there is new code Propagation should be rapid The long lifetimes of these systems mean that users need a way to reprogram them. However, motes are hard to reach, hard to find and numerous. We therefore need to be able to transmit new code into a deployment over a network. However, the lifetime requirements of these systems enforce strict energy budgets. Although we want new code to propagate rapidly into a network, the cost of maintaining a consistent code image must be very low. The loss observed in these networks means that simply propagating code once is insufficient. Transient loss that temporarily disconnects a network should not prevent the network from reprogramming when it reconnects. NEST Retreat, Jan 2004
Relation to Deluge, SPIN, etc. When do you advertise code? How do you suppress control messages? Not a dissemination protocol NEST Retreat, Jan 2004
Solution: Trickle Simple, “polite gossip” algorithm “Every once in a while, broadcast what code you have, unless you’ve heard some other nodes broadcast the same thing.” Behavior (simulation and deployment): Scalability: thousand-fold density changes Maintenance: a few sends per hour Propagation: less than a minute NEST Retreat, Jan 2004
Outline Introduction Trickle algorithm Maintenance Propagation Conclusion NEST Retreat, Jan 2004
Trickle Assumptions Wireless broadcast medium Concise, comparable metadata Given A and B, know which has newer code NEST Retreat, Jan 2004
Idea: Communication As long as each mote communicates with one other, needed updates will be detected In a single cell, one transmission will detect all needed updates Communication is reception or transmission Maintain a communication rate: (receptions + transmissions) <= k NEST Retreat, Jan 2004
Trickle Algorithm Time interval of length t Redundancy constant k (e.g., 1, 2) Pick a time t from [0, t] Maintain a counter c, initialized to zero At time t, broadcast code metadata if c < k Increment c when you hear identical metadata to your own At end of t, pick a new t NEST Retreat, Jan 2004
Example Trickle Execution 1 2 3 t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 t1a 2 3 t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 t1a 2 1 3 t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 t1a 2 1 3 t3a t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 t1a 2 2 3 t3a t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 t1a 2 2 t2a 3 t3a t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 t1a 2 t2a 3 t3a t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 1 t1a 2 t2a t2b 3 1 t3a t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 1 t1a 2 t2a t2b 3 1 t3a t3b t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Example Trickle Execution 1 1 t1a t1b 2 t2a t2b 3 1 t3a t3b t time transmission suppressed transmission reception NEST Retreat, Jan 2004
Outline Problem statement Trickle algorithm Maintenance Propagation Conclusion NEST Retreat, Jan 2004
Maintenance Minimize maintenance cost (transmissions) when there is no new code Keep as close to k as possible Start with three assumptions, relax each Lossless network Perfect synchronization Single-hop NEST Retreat, Jan 2004
Ideal Case k transmissions per interval First k nodes to transmit suppress all others Independent of density NEST Retreat, Jan 2004
Loss NEST Retreat, Jan 2004
Logarithmic Behavior of Loss Transmission increase is due to the probability that one node has not heard n transmissions Example: 10% loss 1 in 10 nodes will not hear one transmission 1 in 100 nodes will not hear two transmissions 1 in 1000 nodes will not hear three, etc. Fundamental bound to maintaining per-interval communication NEST Retreat, Jan 2004
Synchronization NEST Retreat, Jan 2004
Short Listen Effect Some nodes don’t listen much (pick small t values) For example, B transmits three times: t A B C D Time transmission suppressed transmission reception NEST Retreat, Jan 2004
Solution Add a listening period: pick t from [0.5t, t] Listen-only period NEST Retreat, Jan 2004
Effect of Listen Period NEST Retreat, Jan 2004
Multi-Cell Case TOSSIM simulation Logarithmic scaling holds No synchronization, loss from empirical model Nodes uniformly distributed in 50’x50’ area Logarithmic scaling holds NEST Retreat, Jan 2004
Empirical Validation Redundancy: Maté VM implementation (transmissions + receptions) Redundancy: Maté VM implementation - k intervals NEST Retreat, Jan 2004
Outline Problem statement Trickle algorithm Maintenance Propagation Conclusion NEST Retreat, Jan 2004
Choosing Intervals Large interval: low cost, slow to discover Small interval: high cost, quick to discover When there’s new gossip, talk more When there’s nothing new, talk less NEST Retreat, Jan 2004
Speeding Propagation Adjust t: tl, th When t expires, double t up to th When you hear newer metadata, set t to tl When you hear newer code, set t to tl When you hear older metadata, send an update NEST Retreat, Jan 2004
Rate Change Illustration Hear Newer Metadata 2tl th tl th th 2 Time NEST Retreat, Jan 2004
Simulated Propagation k=1, tl=1 second, th=1 minute NEST Retreat, Jan 2004
Empirical Propagation Deployed 19 nodes in office setting Instrumented nodes for accurate time measurements Introduce new code, log installation times k=1, tl=1 second, th=1 minute 40 test runs NEST Retreat, Jan 2004
Network Layout NEST Retreat, Jan 2004
Empirical Results NEST Retreat, Jan 2004
Changing th to 20 minutes k=1 NEST Retreat, Jan 2004
Conclusions Trickle efficiency scales logarithmically with density Can obtain rapid propagation with low maintenance At most 3 sends/hour, propagates in 30 seconds Uses beyond code propagation Changes to data such as routing tables E.g., predicates can scope distance Further examination of tl, th and k needed NEST Retreat, Jan 2004
Questions Tech report available in back NEST Retreat, Jan 2004
NEST Retreat, Jan 2004
NEST Retreat, Jan 2004
NEST Retreat, Jan 2004