Power-aware scheduling

Document Sample
Power-aware scheduling Powered By Docstoc
					courseware




      Power-aware scheduling
                       Jan Madsen

             Informatics and Mathematical Modelling
                 Technical University of Denmark
              Richard Petersens Plads, Building 321
                   DK2800 Lyngby, Denmark
                         Jan@imm.dtu.dk
Mission critical embedded systems

 Based on work by
        J. Liu,
        P.H. Chou,
        N. Bagherzadeh,
        F. Kurdahi
 University of
  California, Irvine
 CODES’01 &
  DAC’01


SoC-MOBINET courseware   Jan Madsen
Mars Rover – Mission

 Perform experiments
      Autonomous mobile vehicle
      Alpha proton X-ray spectrometer
 Imaging
 Travel between different target
  locations




SoC-MOBINET courseware   Jan Madsen
Mars Rover – Conditions

 Surface temperature [-40 oC; -80 oC]
 Communication ~ 11 minute
      No real-time control
      Supervised autonomous control




SoC-MOBINET courseware   Jan Madsen
Mars Rover - System composition

 CPU
      3 images per day
 Motors
      60 cm per min
 Hazard detection
 Heaters
      -80 oC requires
       motors to be heathed




SoC-MOBINET courseware   Jan Madsen
Mars Rover – Power?

 Power sources
      Battery (non-rechargeable)
      Solar panel (free)
 Power consumers
      Digital: imaging, communication, control
      Mechanical: driving, steering
      Thermal: heating motors in the low-temperature
       environment




SoC-MOBINET courseware   Jan Madsen
System-level power manager

 Amdalhs’ law applies to power
      Power savings of a component is scaled to its
       contribution to power usage of the whole system
    If a component draws 2% of the power in a
     system, a 50% power reduction amounts to 1%
     saving to the system
 The power manager must consider all power
  consumers in the entire system and identify the
  major power consumers


SoC-MOBINET courseware   Jan Madsen
System-level power manager

 System-level power consumers
      (Digital) computation domain
            Processors, memory, I/O, ASIC
      Non-computation domains
            Mechanical: motors
            Thermal: heaters
 Major power consumers: mechanical and
  thermal




SoC-MOBINET courseware   Jan Madsen
Power-aware vs. low-power

     Low-power
            Minimize power usage
            Just enough power to meet performance requirement
            No distinction between costly power and free power
            Component-level power managers
     Power-aware
          Best use of available power
               Minimize power usage with low power budget
               Deliver high performance with high power budget
          Distinguish different models of power sources
               Battery, solar, nuclear, etc.
          Track variant power availability
          System-level power managers


SoC-MOBINET courseware    Jan Madsen
Low-power scheduling

 Shutting down subsystems
 Variable-voltage processor scheduling
 Limited applicability to power-aware designs
        Timing constraints are not strongly guaranteed
        Power usage is handled as a by-product
        No tracking to power availability
        No distinction to different energy sources




SoC-MOBINET courseware   Jan Madsen
Low-power scheduling - Example


   p1       r1                             r1
   p2                                                         r1
   p3


                              r1 idle               r1 idle
   p1       r1           r1
   p2                                                              r1
   p3

                                                r1 idle
SoC-MOBINET courseware        Jan Madsen
Power-aware scheduling

 Min/max timing constraints on tasks
      Min timing constraint
            Subsumes precedence as special cases
      Max timing constraint
            Subsumes deadline as special cases
 Min/max power constraints on the system
      Max power constraint
            Total power budget from the available sources
            Hard constraint, must be guaranteed
      Min power
            Free power (solar), minimize power jitter
            soft constraint, best effort




SoC-MOBINET courseware   Jan Madsen
Constraint graph G(V, E)

 Vertices V: tasks                    Edges E: timing
      d(v), execution delay            constraints
      p(v), power consumption            Forward edge: min
      r(v), resource mapping              constraint
                                          Backward edge: max
                                           constraint




SoC-MOBINET courseware   Jan Madsen
Constraint graph G(V, E)

 Schedule                            Timing-valid schedule
      Time assignments to tasks          Timing constraints satisfied
      Finish time                      No resource conflict




SoC-MOBINET courseware   Jan Madsen
Power-aware Gantt chart

     Time view                             Power view
           Bins – tasks                      Power profile
                Horizontal axis – start      Power constraints
                 time, duration
                                              Power properties
                Vertical axis – power
                                                 Spikes, gaps
           Tracks – parallel                    Energy cost
            resources                            Utilization




SoC-MOBINET courseware   Jan Madsen
Mars Rover - Exercise




SoC-MOBINET courseware   Jan Madsen
Mars Rover - Exercise


     Power sources &      Duration     Power @ -40 oC   Power @ -60 oC   Power @ -80 oC
     tasks                (sec.)
     Solar panel                             17              14                11

     Battery pack                          8 max            8 max            8 max

     CPU                  Constant           2                3                4

     Heating two motors      5               8               10               12

     Driving                 10              8                11              14

     Steering                5               4                6                8

     Hazard detection        10              3                4                5




SoC-MOBINET courseware    Jan Madsen
Mars Rover - Solution

      Worst case at –80 oC

 Hd
 St
 Dr
 HW12
 HW34
 HW56
 HS12
 HS34
 CPU


 Power   9    9    16    16    16   16     16   12   18   18   9   9   12   18   18




SoC-MOBINET courseware        Jan Madsen
Power properties
 Power profile P(t)                     Min power utilization
      System-level power
       consumption curve                   (Pmin)
 Power constraints                          Energy utilization from free
      Max power constraint Pmax              sources
            Power Spike: max power
             constraint violation         Energy cost Ec(Pmin)
      Min power constraint Pmin
            Power Gap: min power            Energy drawn from
             constraint violation             expensive (non-free)
 Power-validity                              sources
      A timing-valid schedule with no
       power spikes                       Power-aware trade-off
      Enforce max power budget
                                             Performance  vs. Energy
                                              cost Ec(Pmin)




SoC-MOBINET courseware   Jan Madsen
Mars Rover – Power profile



                                                                  Pmax
 20
                                                                         P(t)


                                                                  Pmin
 10

                            (11 x 75) – (2 x 2 x 10)
            (Pmin) =                                 = 95.2 %
                                       (11 x 75)

                            5x25+5x1+10x7+5x1+10x7
            Ec(Pmin) =                            = 3.4
                                     75


SoC-MOBINET courseware    Jan Madsen
Mars Rover – the real thing!

 Timing constraints



 Three cases w/ different power constraints
      Max power:
           solar + 10W
      Min power
             solar, free
             Best: 14.9W
             Typical: 12W
             Worst: 9W


SoC-MOBINET courseware   Jan Madsen
Scheduling results

     Best case
           Fast, low cost




                                       Worst case
     Typical case                       Slower, high cost
           Slower, increased cost       Same as the existing
                                          serial schedule




SoC-MOBINET courseware   Jan Madsen
Comparisons to schedules

     Existing low-power                   Power-aware schedules
      schedule                               High performance
           Low performance                  High energy cost
           Low energy cost                  Improved utilization of
           Under-utilized free solar         solar power
            power                            Tracks available power
           Does not track power              from different sources
            sources                          Fully constraint-driven by
           Full serialization by hand-       an automated design tool
            crafting




SoC-MOBINET courseware   Jan Madsen
Comparisons in a scenario

     Scenario
           Mission: travel to a target      3 phases: best, typical,
            48 steps away                     worst, 10 min each
     Existing low-power                   Power-aware schedules
      schedule                               Accelerated speed by
           Fixed slow speed                  tracking available power
           Low energy cost in each          Finish earlier before
            phase, but high energy            working in the worst case
            cost in worst case               High performance, low
           Low performance, high             energy cost
            energy cost




SoC-MOBINET courseware   Jan Madsen
Conclusion

 Power-aware design
      Different from low-power
      Deliver high performance by tracking power sources
 Power-aware schedulers
      Incremental scheduling by constraint classification
      Potentials on performance speedup and energy saving
 System-level design tools
      Power manager for the entire system
      Aggressive design space exploration



SoC-MOBINET courseware   Jan Madsen
Incremental scheduling (1)

 (1) Timing scheduling
        Topological traversal of the constraint graph
        Selective serialize tasks that share the same resource
        Prohibit positive cycles
        Proven to find a timing-valid schedule




SoC-MOBINET courseware   Jan Madsen
Incremental scheduling (2)

     (2) Max power scheduling
             Begin with a timing-valid schedule from (1)
             Enforce max power constraint
             Reorder tasks to eliminate power spikes
             Redo (1) for timing violation
             Heuristics applied




SoC-MOBINET courseware   Jan Madsen
Incremental scheduling (3)

     (3) Min power scheduling
             Begin with a power-valid schedule from (2)
             Reorder tasks to reduce power gaps in best-effort
             Deliver same performance with less energy cost
             Heuristics applied
             Results applicable to different constraints




SoC-MOBINET courseware   Jan Madsen

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:9
posted:2/28/2012
language:English
pages:28