PADS Power Aware Distributed Systems Middleware Techniques and Tools

1
PADSPower Aware Distributed SystemsMiddleware
Techniques and Tools

• USC Information Sciences InstituteBrian Schott,
  Bob Parker
• UCLAMani Srivastava
• Rockwell Science CenterCharles Chien

2
PADS Project

• Q How can you extend the dynamic power range of
  sensor networks from quiescent months of
  monitoring to frenetic minutes of activity?
• Architectural Approaches
• Power Aware Research Platform Testbed
• Deployable Power Aware Sensor Platform
• Middleware, Tools, and Techniques
• Power Aware Resource Scheduling in RTOS
• Techniques for Network-Wide Power Management
• Power Aware Algorithms
• Multi-Resolution Distributed Algorithms

UCLA
3
Introduction to UCLA PADS Team

• PI
• Mani Srivastava
• Associate Professor in EE Department / Computer
  Engineering
• Area wireless systems, networked embedded
  systems, low power systems
• Experience Ph.D. Berkeley, 4 years at Bell Labs
• Other faculty
• Rajesh Gupta
• Associate Professor in CS Department at UCI
  (UCLAs sister campus)
• Area embedded and real-time systems,
  architecture, design tools
• Experience Ph.D. Stanford, several years at
  Intel, UIUC-CS
• Students (planned)
• Sung Park, Pavan Kumar, Paleologos Spanos, Vijay
  Raghunathan
• all are Computer Engineering graduate students

4
Planned Research

• Broad goals
• Middleware techniques for JIT power through
  coordinated scheduling and power management of
  computing and communication resources locally at
  a sensor node (RTOS) as well as globally in a
  sensor network (protocols)
• Tools for evaluating and designing the power
  management techniques
• Target 10-30x gain in power efficiency
• Specific subtasks
• Power management within a sensor node
• Power-aware RTOS scheduling under timing
  constraints
• Resource management with energy-speed and
  energy-accuracy control knobs
• Tools for RTOS power management evaluation, and
  power-aware kernel synthesis
• Network-wide power management
• Network resource allocation for global power
  management
• Power-aware network protocols
• Hybrid sensor network simulation framework for
  power vs. quality evaluation of network level
  power management techniques and protocols
• Power management with multimedia sensor data
• Power characterization, extension of PADS
  techniques to streaming multimedia
• Integration with sensor nodes (Rockwell nodes,
  research platform)

5
Power Management Trade-offs in Sensor Networks
Lifetime(power)
Rapidity(latency -1)
Quality(coverage, fidelity)
6
Power-aware Operation

• Intra-node
• hardware circuits
• RTOS software
• Inter-node
• network protocols

7
Power Management in RTOS

• Traditional approaches
• set voltage to match (average) sample rate
• shutdown when idle and wake on demand to exploit
• e.g. predictive approach by Srivastava,
  Chandrakasan, and Brodersen
• Realities of sensor networks
• Latencies are critical
• unlike DSP where only rate (sample period)
  matters
• deadlines important in protocols, target tracking
• Tasks are dynamic
• cannot schedule the tasks statically
• Fortunately, hardware provides many control
  knobs for power-performance trade-off
• CPUs with dynamic frequency/voltage, and shutdown
  mode
• radios with multiple modes and symbol rate
  choices
• Potential for dynamic power management with
  power-quality trade-offs

8
Example Fixed Priority Preemptive CPU Scheduling
in RTOSs

• Consider task set (period, WCET, deadline)
• (10, 3, 10), (14, 7, 14)
• CPU utilization 3/10 7/14 80
• Obvious power management strategies
• Shutdown when idle
• saves 20 power
• Can we slow CPU by 20 ( reduce V) for more
  savings?
• NO, as deadlines will no longer be met
• However, can slow by x 14/13 and lower voltage to
  still meet deadlines, and shutdown during idle
  time
• saves 22.5 in power
• Problem uses WCET (worst case execution time)

9
Variation in Execution Times

• Significant variation in execution time of tasks
• WCETBCET often gtgt 1
• e.g sensor processing time varies depending on
  target activity
• e.g. compressed speech playout has different time
  for talkspurt vs. silence
• e.g. on test run, MPEG decoder time range
  0.003s, 0.15s with average 0.035s

10
Predictive Strategy for Exploiting
Processing-time Variation

• Obvious shut down or reduce voltage if task
  finishes earlier
• Even better predict execution time of task
  instance and dynamically scale voltage even more
  aggressively
• task-specific predictor to exploit history
• significant temporal correlation
• but, some deadlines may be missed!
• leads to packet loss another form of noise!
• Provides power-quality trade-off

11
Example 1 (Simulation)
12
Example 2 (Simulation)
13
Tool Framework for RTOS Power Management
Evaluation

• Goals
• Performance and power modeling of an RTOS
  environment that incorporate
• application-level contracts on power, timing, and
  functionality
• multiple power management policies and scheduling
  disciplines
• Define suitable contractual requirements and
  ways to specify them.
• Evaluate techniques for admission control by
  the RTOS
• Generate RTOS kernels with application-specific
  power management
• Challenges
• timing functionality and power performance are
  interrelated
• power management affects satisfiability of timing
  by scheduling and vice-versa
• define energy-speed and energy-accuracy
  control variables in RTOS to navigate optimal
  scheduling and strategy
• use mathematical programming approaches to
  optimize policy selection and scheduling

14
RTOS Power Analyzer Tool

• Build a power analyzer tool for determines
  power/performance feasibility for a given task
  scenario and application-level service contract
  (related to task timing, functionality, and
  energy budget)
• initial work on bounds on improvement by
  different power management strategies for use by
  a RTOS in a decision procedure
• Inputs
• task description and constraints
• prototype or automatic abstraction from routing
  protocols,
• task deadline, response time, rate and interval
  separate constraints
• a power management policy
• predictive, stochastic, adversarial
• a scheduling discipline
• rate-monotonic, deadline-driven
• a task level power model
• that allows accurate estimation of system-level
  speed versus power points
• Outputs
• satisfiability of timing constraints e.g.,
  deadlines met or missed
• an estimate of power savings

15
Power Management of Radios

• What is given or estimated?
• Tx-Rx distance
• Channel condition
• What does one control?
• Radio modem settings
• transmit power, modulation scheme (type, symbol
  rate, bits/symbol)
• Protocols and their parameters
• frame length, error control (coding, ARQ),
  routing single-hop vs. multi-hop
• What do we get?
• Quality useful bit rate
• depends on raw bit rate, BER, protocol efficiency
• Power computingcommunication energy spent at Tx
  and Rx per useful bit communicated
• Strategy select modem and protocol parameters to
  minimize the power metric for a given quality
  level
• power-quality trade-off
• key problem sender cant wake up receiver

16
Example Power-Rate Trade-off via Modulation

• At 0.001 BER and fixed transmission bandwidth
  withM-ary QAM

17
Example Power-aware Frame Length Adaptation with
TCP
18
Computation vs. Communication Trade-off in Error
Control
Computation Energy
Communication Energy
Code Rate
Code Rate
Total Energy
Lowest energy for a given BER
Code Rate
19
Example Coordinated Adaptation of FEC Frame
Length
Optimal Code Rate
Fixed Code Route
10
)
)
t
t
i
i
14
b
b
9
/
/
J
J
10-2
m
m
12
8
(
(


t
t
8x10-3
i
i
10-4
B
10
B
7
10-3


l
l
u
u
f
f
6
8
e
e
4x10-3
s
s
U
U
5

6

r
r
e
e
4
p
p
10-3


4
10-8
y
y
g
g
3
r
r
10-8
e
2
e
n
n
2
E
E
0
200
400
600
800
1000
1200
1400
0
200
400
600
800
1000
1200
1400
Packet Length (bytes)
Packet Length (bytes)
20
Envisioned MAC-level Power Management Strategy
for Radio

• GIVEN
• - Estimated channel condition and Rx-Tx
  distance- Desired quality (latency, useful
  bits per second)- Radio energy consumption
  characteristics
• SELECT
• - Radio modem parameters modulation scheme,
  symbol rate, bits/symbol, spreading gain,
  transmit power level, carrier frequency-
  Protocol and protocol parameters frame length,
  error control (FEC, ARQ), of hops
• MINIMIZE
• - computation communication energy per
  useful bit

21
Leverage SensIT Research on Protocols

• Known results
• power-aware adaptive link protocols
• optimize computation and communication energy
  spent per good application level bit distributed
  (Joules/bit)
• multihop communication provides power capacity
  benefits
• Traditional multihop routing is power unaware
• focus on topology changes, and metrics such as
  shortest hop, shortest delay, link quality etc.
• power wasted in quiescent state signaling
• using power-based routing metrics, and reactive
  protocols helps
• Our work in SensIT
• coordinated routing and MAC
• large number of collisions with CSMA MAC during
  broadcasts used by routing protocols for probing
• exploit path diversity, and node traffic
  redundancy

22
Example Exploiting Path Diversity Node
Redundancy

• Ideas combine data from different nodes (Data
  Combining Entities), and distribute traffic over
  alternative paths (Spreading)
• increase network lifetime and coverage
• packet disperser and combiner entities
• works with variety of routing approaches
• Evaluation metrics
• time to breakdown
• of depleted nodes
• RMS energy distribution
• Which nodes are important depends on future
  target traffic pattern and user movement
• traditional load balancing is based on only
  present activity
• goal is stochastic lifetime, but practical
  approaches need indirect measures

23
Energy Efficiency Impact
Without load spreading
With load spreading

• An approximate way is a histogram
• area of histogram vs. shape of histogram
• but only approximate (cant average over all
  futures)
• Possible metric to capture the essential
  histogram info
• RMS of the histogram (measures total as well as
  spread)

24
SensorSim Hybrid Simulator

• Motivation study sensor network deployment,
  protocols, applications, and power-quality
  trade-offs at scale in a controlled setting
• Three key capabilities
• Sensor and target modeling
• Target, sensor channel, and sensor transducer
  characteristics
• Power modeling
• Power characterization via data from instrumented
  platforms
• Energy consumer models radio, CPU, sensors
• Energy source models batteries
• Power-quality trade-off analysis and
  visualization
• Hybrid simulation
• selected nodes in a simulation can be real
  nodes
• currently supports only higher layers in real
  nodes
• real applications can run on nodes in a
  simulation
• Current implementation based on ns simulator

25
SensorSim Architecture
app
monitor and control hybrid network (local or
remote)
real sensor apps on virtual sensor nodes
app
GUI
app
socket comm
serial comm
ns
HS Interface
GUI Interface
RS232
Ethernet
gateway
V
V
R
V
Gateway Machine
V
R
modified event scheduler
Proxies for real sensor nodes
Simulation Machine
26
Sensor Node Model in SensorSim
Node Function Model
Micro Sensor Node
Applications
Power Model (Energy Consumers and Providers)
Middleware
State Change
Network Protocol Stack
Radio Model
Sensor Protocol Stack
Network Layer
Sensor Layer
CPU Model
Status Check
MAC Layer
Physical Layer
Sensor 1 Model
Physical Layer
Sensor 2 Model

Wireless Channel
Sensor Channel
27
Battery Model

• Common battery model bucket of constant energy
• Reality delivered energy depends on how the
  battery is discharged
• discharge rate (load current)
• C k/I? where ? up to 0.7
• discharge profile and duty cycle
• operating voltage and power level drained
• Appropriate protocols and power management
  strategy can lead to higher electrical work done
  for the same battery

100
Efficiency
50
0
1
2
3
4
Discharge Current Ratio
28
Radio Model

• Example Values
• Eelec 50nJ/bit
• eamp 100pJ/bit/m2

29
Radio Power Management Example

• Using a 2Mbps WaveLAN NIC model in ns-2
• Dynamic Source Routing (DSR)
• Case 1 node 0 transmits 512 byte packets every
  2s to node 3 for 500s
• Case 2 nodes 0, 6, 1 continually transmit to
  nodes 3, 4, 2
• MAC Layer is responsible for the power control

30
Sample Power Management Strategy
BZR event
Off
BZR event
BZR event
BZR event
Off
BZR event
Transmit
transmit
Receive
Transmit
BZR event
Receive
transmit done
receive done
Idle
transmit
transmit
receive done
Idle
timeout
receive
transmit done
timeout(3 sec)
Sleep
With Power Management
Without Power Management
31
Radio Power Management Simulation
32
Concept Demonstration using Initial Version of
SensorSim

• End User Station
• Location
• Coverage
• Target Reports

Gateway
SensorSim Workstation
Target Detectors