CS294-1 Deeply Embedded Networks Emerging Standards and Course PerspectiveNew Class of ComputingExample usesTypical CharacteristicsThe System ChallengeA Systems ViewWhere Have We Been?Where We’ve been (cont)Impact of radio design?Fresh Look at BlueToothEvents and BufferingMultihop Topology FormationIn-Network ProcessingThroughput & PowerEnergy UsageClosing ObservationsWhat about 802.15.4?Lessons about Methodology1. Compare to a Weak Strawman2. Change many parameters at once and claim the one that you have been talking about accounts for the improvement3. Design for a different load point and compare where yours is intended4. Do very narrow empirical assessment and extrapolate through simulation, neglecting the impact of more general settingSmall Technology, Broad AgendaLarger QuestionsThanksCS294-1 Deeply Embedded NetworksEmerging Standards and Course PerspectiveDavid CullerUniversity of California, Berkeley12/2/0312/2/03 New Class of Computingyearlog (people per computer)streaming informationto/from physical worldNumber CrunchingData Storage productivityinteractiveMainframeMinicomputerWorkstationPCLaptopPDA12/2/03 Example uses•Env. Monitoring, Conservation biology, ...–Precision agriculture, land conservation, ... –built environment comfort & efficiency ... –alarms, security, surveillance, treaty verification ...•Civil Engineering: structures response–condition-based maintenance–disaster management–urban terrain mapping & monitoring•Interactive Environments–context aware computing, non-verbal communication–handicap assistance»home/elder care»asset tracking•Integrated roboticsCENS.ucla.edu12/2/03 Typical Characteristics•# nodes >> # people•sensor/actuator data stream•unattended•inaccessible•prolonged deployment•energy constrained•operate in aggregate•in-network processing is necessary•what they do changes over time=> must be self-organized, self-maintaining and programmed in situ to operate at very low duty cycle12/2/03 The System Challengeapplicationsservicenetworksystemarchitecturedata mgmtMonitoring & Managing Spaces and ThingstechnologyMEMSsensingPowerComm.uRobotsactuateMiniature, low-power connections to the physical worldProcStore12/2/03 A Systems View•Desire for decomposition–Modularity, Optimization, Predictability•Interactions Across Layers / Components–Constraints–Information availability–Control–Performance Characteristics12/2/03 Where Have We Been?•Application-Driven Network Architecture •Emergence of Wireless Sensor/Effector Nodes •Operating Systems for DENs•Low-power MAC, discovery, topology formation •Tools for analysis •Broadcast / Data dissemination •Design Lessons (Lew Girod, Dave Hughes)•Aggregation and in-network processing (Sam)•Multihop Routing for Data Collection (Alec)•Time Synchronization •Ad Hoc Routing•Cluster Formation12/2/03 Where We’ve been (cont)•Directed Diffusion •Localization•Collaborative Signal Processing (Feng Zhao)•Tracking•Multi-resolution Storage•Distributed Control•Coverage•Security•Privacy•Emerging Standards12/2/03 Impact of radio design?•Bluetooth ?•IEEE 802.15.4 ?•Bluetooth and Sensor Networks: A Reality Check–Designed as cable replacement–Connection oriented, frequency hopping–Narrow interface–TDM - master/slave(7) piconet–ScatterNet ???12/2/03 Fresh Look at BlueTooth•Fixed MAC•Similar event-driven•Discovery and connection mgmt below HCI–Master: inquiry–Slave: inquiry scan•Pre-established connection above•Build self-assembly out of connection ???•Timing and pwr mgmt invisible to appln–App adapts to radio, not reverse•Very small stack –10% of Bluetooth–3 KB for UART HCI–MAC in HW12/2/03 Events and Buffering•Retain buffer swap•Start/End ptr across layers to allow encapsulation12/2/03 Multihop Topology Formation•Two radios per node–Master (connects to children)–Slave (connects to parent)•Grow Tree as very slow flood–Turn on slave and look for master–If success, turn on M and allow children•Backtrack on fail–Try alternating parent–3 connect-fail on node with 7 children»Disconnect child»On disconnect, disconnect all children–Convergence???–Maintenance over time?•How well connected can network be?SMSMSMSM12/2/03 In-Network Processing•Query Graph subset of radio graph•Devices can negotiate “sniff mode”–App has no control over timing•App adapts to TDM–Pipelines query processing across epochs–Cannot go into deep sleep•No rate adaptation to contention•No snooping12/2/03 Throughput & Power•Small fraction of peak•5x ChipCon & RFM•Similar uJ/bit•High Power–30 mW radio on–39 mW inquirable–89 mW waiting for conn–136 mW maintain conn–200 mW @ 6 KB/s•Sniff save 5 mW12/2/03 Energy Usage•Radio off => rediscover•10-30 sec to discoverMicaWorstcase12/2/03 Closing Observations•Scatter nets still not supported•Certification•Applns denied relevant information•Connection maintenance expensive•Sniff not much value•Perhaps this all changes with “single chip” version–MCU shared with application12/2/03 What about 802.15.4?•Sensor nets at least a secondary goal–Game controllers primary•Direct Sequence Spread Spectrum –instead of freq. Hop (O-QPSK)•Phy & MAC separate from network (Zigbee)•Simple, controllable MAC•Attention to low duty-cycle devices•0-104 byte packets12/2/03 Lessons about Methodology•Emerging area, so not just X improvement over established basis•Reaching beyond current technology•Define method of evaluation along with new idea•Still, some classic pitfalls to avoid12/2/03 1. Compare to a Weak Strawman•My fancy routing protocol is 40% better than all nodes flooding (when there is only one or two “communications” going on at a time).•My fancy scheduling algorithm uses half the energy of naively burning full time (when there may be simple ways of reducing energy cost of most prevalent state).•This piece of steel is a million times stronger than that wet noodle.•Weak strawthings are for negative results–This piece of lasagna is ONLY 40 stronger than that wet noodle–If it is all you’ve got, estimate optimal so you can see the spread12/2/03 2. Change many parameters at once and claim the one that you have been talking about accounts for the improvement•My adaptive clustering protocol (which happens to also compress n values to 1 at each hop) is 40% better than tree-based
View Full Document
Unlocking...