Power Aware Software Architecture

1
Power Aware Software Architecture

• Rajesh K. Gupta
• University of California, San Diego

Cristiano Pereira, Ravindra Jejurikar, Yuvraj
Agrawal, Manjari Chachharia, Sandeep Shukla,
Mukesh Rajan
In collaboration with Sandy Irani, Mani Srivastava
2
Computing In New Spaces

• Generational shift in computing devices
• lot more of everything including networking and
  communications
• lot less of power, energy, volume, weight,
  patience
• Application is everything, the possibilities are
  limitless
• System architectures are due for an overhaul
• the architectures are (radically)
  changed/challenged
• the programming context is changed
• the system software contract is changed
• new awareness location, power, timing,
  reactivity, stability

power
3
Outline

• The case for power awareness in
• application development
• system software
• Managing power in the OS
• knobs and strategies
• Making software power aware
• the hardware knobs (DVS, DPM)
• the application knobs (duty cycling, criticality,
  aesthetics)
• An ongoing experiment

4
The Case for Power Awareness

• Limited availability
• Energy and power uses of new devices is markedly
  different from laptops and notebook computers
• much wider dynamic range of power demand
• increasing share of memory, communication and
  signal processing
• multiple power use modalities depending upon
  application
• immortal, paging-mode RX, lifeline TX,
  mission-mode

5
Power Management Places

• Hardware firmware
• many techniques for low power design in circuits,
  architectures, etc.
• but dont know the global state and
  application-specific knowledge
• Users
• dont know component characteristics, and cant
  make frequent decisions
• Applications
• operate independently
• and the OS hides machine information from them
• OS (role in resource allocation and sharing)
• it is a logical place for dynamic power
  management
• early results show 50-70 savings due to f/v
  scaling Weiser94
• application-specific constraints and
  opportunities for saving energy that can be known
  only at that level

6
Operating System Directed Power Management

• Significant opportunities in power management lie
  with application-specific knobs
• quality of service, timing criticality of various
  functions
• Needs of applications are driving force for OS
  power management functions power-based API
• collaboration between applications and the OS in
  setting energy use policy
• OS helps resolve conflicts and promote
  cooperation
• OS is the most reasonable place, but
• OS should incorporate application information in
  power management
• OS should expose power state and events to
  applications for them to adapt.

7
Power Savings Mechanisms
A

• Dynamic Power Management (DPM)
• When a device is idle, it can transition to
  low-power sleep states.
• Current trend is to design devices with multiple
  sleep states and provide device driver hooks to
  change these states under OS control.
• Dynamic Voltage Scaling (DVS)
• A device can be run at different speeds at
  different power levels
• Execution of jobs can be slowed down to save
  power as long as all jobs are completed by their
  deadline.
• Application level knobs
• quality and performance measures, application
  tolerances

B
C
8
A. Dynamic Power Management

• When a device becomes idle, it can transition to
  lower power usage state.
• A fixed amount of additional time and energy are
  required to transition back to active state when
  a new request for service arrives.
• What is the best time threshold to transition to
  the sleep state?
• Too soon pay start-up cost too frequently.
• Too late spend too much time in the high-power
  state
• Generally, transition to sleep state when the
  cost of being in active state is at least the
  cost of waking up.

A
9
Our Work In This Context

• We have developed quantitative bounds on the
  quality of DPM algorithms based on Competitive
  Analysis TCAD 01
• provides a basis for DPM strategy comparison
• Developed DPM strategies for devices with both
  multiple active and multiple sleep states TCAD
  02
• Design and analyze algorithms for systems that
  allow both DPM and DVS SODA 03, TECS02
• Important conclusions
• Not all power states are useful in a given DPM
  strategy
• DPM generally useful for improving quality
  measures.

A
10
Competitive Analysis

• Deterministic algorithm (ski rental)
• Transition to sleep state when the cost of being
  in active state is at least the cost of waking
  up.
• Normalize cost of transitioning from sleep to
  active state to 1.
• Power consumption rate of active state is ?.
• This algorithm is 2-competitive.
• 2 is the best possible competitive ratio for any
  deterministic algorithm.
• Probabilistic algorithm
• Idle period length generated by known
  distribution with density function p(t).
• Choose threshold T to minimize cost
• For any distribution p(t), the expected cost of
  the above algorithm is within e/(e-1) of the
  optimal cost. Furthermore, there is a
  distribution for which no algorithm can be better
  than e/(e-1) times optimal.

A
11
Multi-state DPM Case

• Let there be k1 states
• Let State k be the shut-down state and 0 be the
  active state
• Let ?i be the energy dissipation rate at state i
• Let ?i be the total energy dissipated to move
  back to State 0
• States are ordered such that ?i1 ? ?i
• ?k 0 and ?0 0 (without loss of generality).
• Power down energy cost can be incorporated in the
  power up cost for analysis (if additive).
• Now formulate an optimization problem to
  determine the state transition thresholds.

A
12
Lower Envelope Idea
State1
State2
State3
State 4
Energy
For each state i, plot
Time
t1
t2
t3

• LEA can be deterministic or probabilistic
• PLEA is e/(e-1) competitive.

A
13
Power-Latency Tradeoff

• Tasks arrive through time and take time to run
• If the device is busy when a task arrives, it
  waits in a queue
• Idle period begins when device finishes current
  job and the queue is empty
• If device transitions to sleep state in an idle
  period, some latency is incurred as device
  transitions to active state.
• This in turn effects (shortens) the length of
  future idle periods.
• Power-Latency tradeoff extremes
• Minimize latency always stay in the active
  state.
• Minimize energy usage delay completing any tasks
  until they have all arrived.

A
14
Experimental Study IBM Mobile Hard Drive
Trace data with arrival times of disk accesses
from Auspex file server archive.
A
15
IBM Mobile Hard Drive
A
16
B. Dynamic Voltage Scaling

• Device which can run at any speed s.
• Power consumed if running in state s is given by
  convex function P(s).
• Jobs arrive through time. Job j has
• Arrival time aj
• Deadline bj
• Work required Rj
• Schedule S (s, job)
• s(t) is the speed of the device at time t.
• job(t) is which job is executed at time t.

B
17
Dynamic Voltage Scaling(Dynamic Voltage Scaling
- No Sleep DVS-NS)

• Schedule S is feasible for set of jobs J if for
  every j in J
• Cost of Schedule S is

B
18
DVS with Sleep State (DVS-S)

• Schedule S ( s, job, h )
• h(t) sleep or on
• If h(t) sleep, then s(t) 0.
• Power is a function of speed and state
• P(s, state) P(s) if state on.
• P(s, state) 0 if state sleep.
• P(0) ? is power required to keep device active
  with no tasks running.
• Let k be the number of times the device
  transitions from sleep state to the on state
• Cost of a schedule S is

B
19
Critical Speed

• If the cost to transition from sleep state to the
  on state were 0, the optimal speed for all jobs
  would be the s that minimizes (Rj/s) P(s)
• This is the s that satisfies P(s) s P(s).
• Call this Scrit, the critical speed for ?.
• If we compress the execution of a task by x,
• we expend additional energy because we execute
  the job faster
• we save ? x.
• Scrit is the point at which it is no longer
  beneficial to compress the execution of a task.
• Our approach
• Decide on Active/Idle intervals (determined by
  critical speed)
• Decide on Sleep/On intervals (determined by the
  cost of staying on)

B
20
Implementing DVS

• Often done using slowdown factors
• can be static or dynamic
• For example
• Given a frequency range of fmin ,fmax
• Slowdown factor is frequency scaled to ?min,1,
  where ?min fmin /fmax..
• When we use a slowdown factor of ?, we set the
  frequency to, f ? fmax .
• The voltage is changed to the minimum voltage
  supported at f.

B
21
Slowdown Factors

• Much of the work on slowdown factors has been in
  the context of real-time systems
• makes sense since we need something to tradeoff
  against the power saved
• Known results
• Essentially use schedulability tests to determine
  the amount of slowdown possible
• Along with the attendant assumptions and almost a
  repeat of Real Time research history...

B
22
C. Enable Application Knobs

• Need API Provide ways by which Application, OS
  and Hardware can exchange energy/power and
  performance related information efficiently.
• Need Middleware Facilitate a continuous dialogue
  / adaptation between OS / Applications.
• Need HAL Facilitate the implementation of power
  aware OS services by providing a software
  interface to low power devices

C
23
Power-aware API Requirements

• Independent of Hardware and RTOS implementations
• enables its use in different hardware platforms
• for this all routines should access the HAL
  (Hardware Abstraction Layer) rather than the
  Hardware directly
• enables its use in different RTOS as well as its
  use with different scheduling strategies
• do not count on specific RTOS info and/or
  specific schedulers
• Services provided
• processor frequency scaling and low-power state
  transitions
• with costs of making such transitions
• battery status (if the system is battery based)
• appropriate routines to control energy-speed and
  energy-accuracy knobs available on I/O devices
• network interface, serial interface, LCD, etc.

C
24
Power-aware API

• The applications interface provides the following
  services
• The application is able to
• tell RT information to OS (period, deadlines,
  WCET, hardness)
• create new threads
• tell OS time predicted to finish a given task
  instance
• depending on the conditions of the environment
  (application dependent and not yet implemented)
• OS must be able to predict and tell applications
  the time estimated to finish the task
• depends on the scheduling scheme used
• A hard task must be killed if its deadline is
  missed
• matter of policy in the context of application
  use.

C
25
A Power-Aware Software Architecture
C
26
Power Aware Software Architecture

• PA-API (Power Aware API)
• interfaces applications and OS making the power
  aware OS services available to the application
  writer.
• PA-OSL (Power Aware Operating System Layer)
• implements modified OS services and active
  components such as a DPM manager.
• PA-HAL (Power Aware Hardware Abstraction Layer)
• interfaces OS and Hardware making the power
  control knobs available to the OS programmer.

C
27
Software Architecture

• PA-API - Power aware function calls available to
  the application writer.
• Some functions of this layer are specific to
  certain scheduling techniques.
• PA-Middleware - Power aware services
• implemented on the top of the OS (power
  management threads, data handling, etc...).
• POSIX - Standard interface for OS system calls.
• This isolates PA-API and PA-Middleware from OS.
• PA-OSL - Power aware OS layer.
• Calls related to modified OS services should go
  through this level. Also isolates OS from PA-API
  and PA-Middleware.
• PA-HAL - Power Aware Hardware Abstraction Layer.
• Isolates OS from underlying power aware hardware.
  
• Modified OS services
• Implementation / modification of OS services in a
  power related fashion. Ex scheduler, memory
  manager, I/O, etc.

C
28
Layer Functionality
C
29
DVS Related Functions

• paapi_dvs_create_thread_type(),
  paapi_dvs_create_thread_instance()
• creates type and instance of a task respectively
• paapi_dvs_app_started(), paapi_dvs_app_done()
• delimits execution of useful work in a thread.
  Tell the OS whether the task has finished
  execution or not.
• paapi_dvs_get_time_prediction(),
  paapi_dvs_set_time_prediction()
• get current execution time prediction for a given
  thread
• paapi_dvs_set_adaptive_param()
• set the paremeters of the adaptive policy (it
  will be described later) for a given task.
• paapi_dvs_set_policy()
• choses the policy to be using for DVS

C
30
DVS Related Functions (contd.)

• paosl_dvs_create_task_type_entry(), ...
• create a type and an instance of a thread in the
  kernel internal tables of type and instance
  respectively
• paosl_dvs_killer_thread()
• kills a thread that missed a deadline
• pahal_dvs_initialize_processor_pm()
• initialize structures for processor power
  management
• pahal_dvs_get_current_frequency(),
  pahal_dvs_set_frequency_and_voltage()
  pahal_dvs_pre_set_frequency_and_voltage(),
  pahal_dvs_get_frequency_levels_info()
  pahal_dvs_post_set_frequency_and_voltage()
• functions to switch processor among possible
  frequencies levels
• pahal_dvs_get_lowpower_states_info(),
  pahal_dvs_set_lowpower_state()
• functions to switch processor among low power
  states

C
31
DPM Functions

• paapi_dpm_register_device()
• just register the device to be power managed
• paosl_dpm_deamon()
• implements the actual policy for a specific
  device. This deamon uses PA-HAL functions to
  decide on how to switch devices among all
  possible states.
• pahal_dpm_device_switch_state()
• switch devices state
• pahal_dpm_device_check_activity()
• check whether the device has been idle and for
  how long. This functions needs support from the
  device driver.
• pahal_dpm_device_get_info(), pahal_dpm_device_get
  _curr_state()
• gets information about the device and about its
  current state respectively
• Others
• functions for helping implementing power
  policies. For example
• pahal_battery_get_info() gets battery status

C
32
Current Status

• API specification available from
• http//www.ics.uci.edu/cpereira/pads/
• Implementation
• eCOS RTOS
• open source, Object oriented and highly
  configurable RTOS (by means of scripting
  language)
• Hardware platforms we are currently working with
• Linux-synthetic (emulation of eCos over Linux -
  debugging purposes only)
• Compaq iPaq Pocket PC - StrongARM SA1110 based
  platform
• Accelent IDP (Integrated Development Environment)
  - also StrongARM SA1110 based.
• LRH Intel evaluation board 80200EVB - Intel
  Xscale based

33
DPM Algorithms Implemented

• A predictive RMS low-power scheduling
• It validates the power-aware API implementation
• assumes periodic tasks and deadline period
• The predictive scheduler implementation is
  divided as follows
• tables and variables manipulation
• admission control and static slow down factor
• dynamic slow down factor computation (time
  prediction)
• deadline management (hard deadline tasks)
• The processor frequency and voltage are scaled
  according to the time predicted by the OS
• The application can also predict the execution
  time in order to enhance accuracy.

C
34
Experiments - XScale Processor
For varying voltage
All measurements executing a busy loop
35
Using Power Aware OS Example

• The scheduler adapts frequency according to the
  real time parameters passed in as parameter on
  the thread type.
• The frequency is adjusted by means of slowdown
  factors (a factor can also speed up the processor
  if it is gt 1).

deadline

• void main()
• mpeg_decoding_t
• paapi_dvs_create_thread_type(100,30,100,hard)
  
• paapi_dvs_set_policy(SHUTDOWN STATIC
• DYNAMIC ADAPTIVE)
• paapi_dvs_create_thread_instance(
• mpeg_decoding_t, mpeg_decode_thread)
• ...

WCET
period
void mpeg_decode_thread() for ()
paapi_dvs_app_started() / original code /
mpeg_frame_decode() paapi_dvs_app_done()

Selects the DVS policy for all threads
Kills the thread instance when deadline is missed
C
36
An Experiment

• Application OS running on 80200 XScale board
• Altera FPGA board generating interrupts to wake
  up the processor
• Maxim board providing voltage scaling
• Host PC for debugging and for loading the App.
  OS into the board

37
The Experiment with DVS

• Shutdown when idle
• as soon as CPU becomes idle shutdown the
  processor
• Shutdown static slow down factors
• offline slow down factors are applied. The CPU is
  shutdown when idle.
• Shutdown static slow down dynamic slow down
• run-time slow down factors are computed based on
  a history of execution times in addition to the
  static and shutdown
• Shutdown static slow down dynamic slow down
  adaptive slow down
• a deadline driven factor is also applied in
  addition to the other factors and shutdown. This
  factor adapts itself according to number of
  deadline missed in a previous window of
  executions.

38
DVS Experiment

• Four parameters are defined for the adaptive
  factor
• of deadlines missed tolerable (D) every W
  executions
• Window size (W)
• Lower bound for the factor (L)
• Increments and decrement steps (Inc and Dec)
• For every W executions
• if the number of deadlines missed is less than D
• lower the adaptive factor by Dec if it is greater
  than L, otherwise keep it as it was.
• if the number of deadlines is greater than D
• increment the adaptive factor by Inc.

39
Application Set

• Three different real applications running
  concurrently
• An MPEG2 decoder
• An ADPCM (Adaptive Differential Pulse Code
  Modulation) speech enconding
• Floating point FFT application

40
Task Set

• We used three tasksets based on the applications
  described earlier as shown in the table below

41
Frequency Voltage Scaling

• For the 4 schemes and the 3 tasksets experimented
  we measured processor power consumption using a
  shunt resistor and a DAQ board.
• The voltage of the Xscale processor is
  dynamically varied according to the frequency as
  in the table below

42
Results Taskset A

• Column deadlines missed shows the number of
  deadlines missed per task (T1, T3, T4) for a
  total of 415/207/138 executions respectively. For
  the adaptive algorithm, M varies as the number
  between parentheses, Inc0.1, Dec0.5, W10 and
  D20

43
Results Taskset B

• Column deadlines missed shows the number of
  deadlines missed per task (T2, T3, T4) for a
  total of 130/65/43 executions respectively

44
Results Taskset C

• Column deadlines missed shows the number of
  deadlines missed per task (T1, T3, T5) for a
  total of 130/65/43 executions respectively

45
OS-directed DVS Results
46
Using Application-level knob

• Example Image Compression Algorithm
• tradeoff image quality against energy available
  by varying the compression parameters such as BPP
  (bits per pixel)
• The image compression algorithm is ran in a
  continuous loop with battery polling every 10
  secs.
• A simple power tradeoff policy is added to adapt
  the quality of the image against the battery
  voltage left.
• Whenever the battery drops 30mV the application
  adjusts the image BPP by -0.5 starting at 1.5.
• For a cut-off of 4020mV, the battery life is
  extended from 290 seconds to 340 seconds.

47

• The battery life is extended by 18 with a slight
  ( not noticeable by human eye) degradation of
  image quality

48
Concluding Remarks

• Computers with radios present a very wide range
  of system optimization opportunities for power
  performance
• Efficient power and energy management is key to
  enabling new range of applications
• Energy efficiency is a system-level concern that
  cuts across components, functionality layers and
  implementations
• Application programming needs to be energy aware
  and provide knobs for the system designer to
  incorporate in DPM.

49
Yes, but Microsoft...

• what are they doing?