EnergyCentric Scheduling for RealTime Systems

1
Energy-Centric Scheduling for Real-Time Systems

• Prof. Jan Madsen
• Informatics and Mathematical Modelling
• Technical University of Denmark
• Richard Petersens Plads, Building 321
• DK2800 Lyngby, Denmark

2
Outline

• The need for low power
• Design of real-time systems
• Power-aware design

3
Towards Ambient Intelligence Weiser

• Wireless network delivers infotainment,
  communication, navigation, ... anyplace, anytime,
  for every citizen ...
• Hidden, pervasive computing. IT to background,
  people in the foreground, improves quality of
  life in non-invasive way ...
• Things see, listen, feel, becomes sensitive and
  adaptive to people ...

4
Electronic Devices Support Athletes
ECG, Blood Pressure
Blood Composition (e.g. lactate)
Wearable Digital Assistant
Wireless Link to Coach and Med Team
Multiple Hop BAN
Position Force Sensors
curtsies Rudy Lauwereins (MPSOC02)
5
Smartshirt - wearable computing
6
... or implants
7
Electronic devices for diagnostics
8
Smart pills 1st generation
9
Smart pills 2nd generation
10
Global System for Ambient Intelligence

• Multimedia, games
• QoS
• GPS
• Global connectivity
• Biometric input
• Health ...
• Ambient control

10 ... 100 Gops 0.1-2W
11
Global System for Ambient Intelligence

• SoC
• Wearable Assistants
• Multimedia, games
• QoS
• GPS
• Global connectivity
• Biometric input
• Health ...
• Ambient control

RF
See Hear Feel
Speak Show Stimulate
IF
IF
10 ... 100 Gops 0.1-2W
gt100/person aura
after Rudy Lauwereins (MPSOC02)
12
What are the properties of these Ambient
Intelligence architectures

• Transducer node
• Ultra low energy (100Mops/mW)
• Low flexibility
• Ultra low cost (1)
• 1..10 Mtr (small size)
• Low clock frequency
• DSP and RF dominated
• Chip package codesign
• Ultra fast hardware design

• Assistant node
• Low energy (10..50 Mops/mW)
• High flexibility
• Low cost (100)
• 10..100 Gops, gt100 Mtr
• High clock frequency
• Data-intensive, dynamic tasks
• Task and data concurrency
• Incremental software design

PLATFORM
PACKAGE in a week
_at_ 100..1000 times power efficiency of todays µP
13
Design challenge
min
Design cycle
14
Design of real-time systems
15
Principles of mapping
Partitioning/clustering
Allocation
Mapping
Scheduling
Communication
1
3
2
16
Power consumption

• PCMOS Pstatic Pdynamic
• Pdynamic a f C Vdd2
• Power minimization, lower
• switching activity
• clock frequency
• capacitive load
• supply voltage

17
Power reduction

• Dynamic power management (DPM)
• Based on processor power modes
• Intel 80200
• Dynamic voltage scaling (DVS)
• Frequency and supply voltage can be adjusted at
  run-time
• Usually these are discrete values and not
  continuous

18
Power reduction DPM
r1
r1
r1
19
Power reduction DVS
Power profile
1
2
3
a
mem
1
2
3
20
Power reduction DVS
21
Optimizing a single task
Exploring the design space
22
Optimizing a single task
Exploring the design space
f(t,pe)
Mapping
time
23
Optimizing a single task using DVS
Exploring the design space
24
Optimizing a single task using DVS
25
Optimizing three tasks
t1
26
Optimizing three tasks
t1
p2
2
p1
Mapping
3
27
Optimizing three tasks
t1
p2
2
p1
Mapping
3
28
Optimizing three tasks
t1
p2
p1
Mapping
3
29
Contributions by Flavius Gruian

• Task level scheduling
• Power optimization of a single task
• Task group scheduling
• With and without dependencies
• Uni- and multi-processor systems
• Architecture selection and scheduling
• Considering task assignment as part of the
  optimization

30
Task level scheduling
31
Task group scheduling

• Tasks with dependencies (task graph)
• Static scheduling on uni- and multi-processor
  systems
• List scheduling with energy-sensitive priority
  function
• Tasks without dependencies (task set)
• Dynamic scheduling on uni-processor systems
• Minimum required task speed of RM and EDF
  scheduling
• Extending EDF to include an ordering policy for
  achiving low power at run-time

32
Task graph scheduling
33
Task set scheduling

• Maximum required speed
• Method to achieve the smallest possible
  processing speeds for which the task set is still
  schedulable.
• This part is done off-line
• Applied to both EDF and RM scheduling
• Slack distribution
• An early finishing task may pass unused processor
  time to the next tasks
• Applied to RM scheduling

34
RM with slack distribution
Post-execution analysis, tasks stretch to max
processor utilization
As All, but assuming exact knowledge of task exec
pattern
Run-time speed scheduling using slack
distribution and stochastic scheduling, plus
off-line and strech
Off-line RM-MRS followed bysimple run-time speed
schedule
35
Task set scheduling

• Uncertainty-based scheduling
• Probabilistic execution times of tasks
• Off-line task sequencing to improve run-time
  decisions
• Run-time speed scheduler which adjust processor
  speed when ever a task finishes
• UBS for EDF

36
Architecture selection and scheduling

• Solving the combined task assignment and
  scheduling problem
• Solution for fixed-speed and variable-speed
  processors