UltraLow Power Data Storage for Sensors

1
Ultra-Low Power Data Storage for Sensors Gaurav
Mathur Peter Desnoyers Deepak Ganesan Prashant
Shenoy IPSN'06
2
Motivation

• WSNs
• Local storage is required in a lot of WSN
  applications
• Sensor nodes have limited capabilities
• Current storage subsystems on sensor network
  platforms do not exploit technology trends

• Flash memories
• High capacity
• Energy efficiency
• Low price

GOAL Minimize energy consumption
Communication VS Computation VS Storage Is there
any energy-efficient platform for Sensor
Networks? What is the relation between storage
and communication energy cost?
3
Overview and Contributions

• Examine available flash-based storage options for
  sensor platforms and find the most
  energy-efficient for sensor networks
• Parallel NAND flash
• Compare storage energy cost with communication
  cost
• Storage cost can be two orders of magitude less
  than communication cost
• Examine implications for Sensor Network design,
  evaluate impact and quantify energy reduction
  achieved sensor network services
• Reduction of one order of magitude for
  communication, data aggregation, and lower cost
  for localization.

4
Questions?
5
Introduction

• Wireless Sensor Networks area of significant
  research
• Limited resources
• Optimizations are needed to ensure long lifetime
• Computation vs Communication trade-off
  influenced
• Algorithm design
• Sensor network platform design

6
What about storage?

• Storage subsystem has undergone little change
• Mica motes provide limited storage (lt1MB)
• Energy cost equivalent to or greater than that
  of communication.
• What is the cost of storage?
• Is it rational to use in-network storage
  techniques?
• If storage consumes more energy than
  communication -gt focus on centralized
  data-collection systems
• If communication requires more energy than
  storage then storage techniques should be
  exploited.

7
Understanding trade-offs

• What is the most energy-efficient storage
  platform for sensor devices ?
• How does the energy cost of storage compare to
  that of computation and communication ?
• What are the implications of an ultra-low power
  storage subsystem on sensor net design ?

8
Methodology

• Find which is the most energy-efficient and
  flash-based storage device for sensor networks
• Measure active and sleep-mode energy consumption
• Compare communication, computation and storage
  costs
• Examine energy reduction for energy-efficient
  local storage for services
• Communication
• Data aggregation

9
Flash Memory

• Flash memory is suitable for Sensor Network
  devices
• low energy consumption
• ultra-low idle current
• high capacity
• Available as component for circuit assembly or as
  standardized removable device (MMC, SD).
• Interfaces for flash devices
• Serial transfering one bit at a time
• Parallel transfering one byte at a time (8 bits)

10
Flash Memory contd.

• MMCs
• translate the parallel NAND interface to serial.
• microcontroller for erasure, page remapping,
  ECC, wear leveling,
• ()simplifies system design
• (-) increases power consumption due to the
  additional internal circuitry.
• Surface-mount NAND devices
• ()Eliminate above overhead
• Memory managment performed in SW or HW

11
Evaluation of Flash Devices

• Serial NOR
• Atmel 512KB used on Mica motes and STM TelosB
• MMC
• Designed and fabricated MMC adapter for Mica
  series
• Drivers in TinyOS
• Tested four MMC devices and report the results
  for the best performing one (Hitachi MMC)
• NAND flash
• Designed and built parallel NAND flash board
• Drivers in TinyOS
• Devices Toshiba, Micron

12
Evaluation of Flash Devices

• Measure power consumption in
• active mode when performing reads/writes/erasur
  es
• sleep mode current drawn by the device in its
  lowest power
• Measurement of all devices was performed on a
  Mica2 mote.

13
Evaluation of Flash Devices - Results
14
Evaluation of Flash Devices - Results

• Total Energy Consumption for Flash and CPU
• Perform ECC for NAND flash in software
• Four times the energy consumption of the flash
• Special-purpose hardware can reduce overhead
• Data transfer cost from RAM to Flash (software
  implementation)
• Can be reduced using hardware support(access to
  SPI port, DMA controller)

15
Evaluation of Flash Devices - Summary

• Parallel NAND flash is 21 times more efficient
  than Telos STM flash and 407 times Mica2 Atmel
  flash.
• MMC is based on NAND technology BUT internal
  microcontroller increases idle current as well as
  energy consumption for reads/writes/erasures
• Byte-wide interface of parallel NAND
• () significantly faster, more efficient than
  bit-serial ones.
• (-) supporting large number of I/O
  pins of parallel interface may be difficult on
  low power embedded systems.
• An ideal storage solution
• combine the performance of the parallel NAND
  flash with the lower pin count of serial
  interfaces.
• Platform-level optimizations needed
• Hardware support for data transfer reduce ECC
  overheads

16
Comparison of computation, communication and
storage costs

• Research in wireless sensor systems was focused
  on the trade-off between computation and
  communication. Storage was ignored.
• Computation lt storage ltlt communication
• Challenges conventional wisdom of trade-offs
• Example Seismic monitoring application
• optimized to last for two years running on the
  MicaZ.
• generates 512 bytes of data/sec, stores them on
  NAND flash storage,
• The lifetime of each sensor would reduce by
  only 6 weeks, having stored 28GB of data.

17
Impact of Storage in WSN Applications

• In-network Query Processing
• Existing sensor network deployments
• centralized data collection for query processing
• In-network storage not exploited
• With parallel NAND flash
• Higher degree of local data archival and indexing
  for in-network query mechanisms
• Use of history for efficient network-level
  compression
• More history More accurate models
• Network-level Compression
• In-network data aggregation schemes rely on
  hashtables to perform duplicate packet
  suppression.
• Hashtables (too large for RAM), can be
  constructed on flash storage.
• Flash-based data management schemes (MicroHash)
• Custody Transfer for Delay Tolerant Networks

18
Impact of Flash Storage on Communication Costs

• GOAL Minimize the number of times the radio is
  power cycled (radio startup and shutdown costs
  are high!)
• BMAC uses a per-packet preamble
• high per-packet transmission cost
• Efficient local storage allows
• usage of simple batching mechanisms
• amortize radio startup and shutdown energy costs
  over a larger number of data bytes.
• BUT increased latency of data collection
• Smaller percentage of duty cycling means larger
  preamble
• Batching Approach reduces communication energy
  costs up to 58x

19
Impact of Flash Storage on Data Aggregation

• High-capacity energy-efficient local storage
  allows
• larger amounts of data to be accumulated and
  compressed at once
• more efficient compression leading to lower
  transmission costs.
• Lossless Compression (Huffman encoding)
  computationally expensive, better when amount of
  data increases
• Lossy Compression low computational cost, high
  compression ratio
• Feature Extraction or Event Detection low
  complexity

Assume 60N common computational complexity for
all Schemes
20
Conclusions

• Parallel NAND flash the most energy efficient
  storage device for sensor networks.
• 100-fold more energy efficient than serial NOR
  flash
• Storage energy reduction motivates re-examination
  of computation-communication-storage trade-offs,
• Storage cost is two orders of magnitude less
  than for communication
• Evaluate impact on Sensor Network design
• One order of magitude reduction for communication
  and data aggregation.