ben

Document Sample
ben Powered By Docstoc
					          ECE 555
Real-Time Embedded Systems

   Real-Time Information
      Dissemination

        Presented by
         Ben Taylor
                             1
                  Outline
• Introduction
  – What is information dissemination?
• Solutions
  – System model
  – Feedback Control Theory Solutions
• Results and Performance
• Summary

                                         2
What is Information Dissemination?
• Publishers and Consumers of information,
  known as subscribers
  – Specify constraints on data, metadata
  – High subscriber count
     • Sensor networks, surveillance systems, etc
• Controlled response time
  – Information is valuable in a specific time period
• Valuable Information at the Right Time (VIRT)


                                                        3
           Metadata Matching
• Metadata matching of constraints
  – Can’t reevaluation all subscriptions in each control
    period
  – High-Priority tasks reevaluated within bounded
    response time
  – Number of low priority tasks maximized, QoS
• Cost to evaluate a subscription varies at runtime
  – Changing number of publishers and consumers
  – Complexity of constraint
  – Unpredictable update arrival time
• How to achieve bounded response time?
                                                           4
                Feedback Controller

Set point for
average
response time         Job budget in kth
                      control period




                                          5
             System Modeling
• System identification approach
  – r(k) = Σair(k-i) + Σbin(k-i)
  – i from 1 to na and nb
• Use Least Squared Method with white
  noise to validate models
  – na = 0
  – nb = 1
• System model r(k) = b1n(k-1)
                                        6
          Root-Locus Design
• A PI controller
  – Integral is used to help eliminate steady-state
    error
  – No derivative because it can amplify noise
• In the Z-domain F(z) = K1(z-K2)/(z-1)
  – K1 = 1 / b1
  – K2 = 0
• G(z) = z-1
                                                      7
              Model Variation
• System model is not
  perfect. Need to
  handle variation
• The system model
  is approximately
  linear between response time and subscription
  reevaluation.
• Model system as r(k) = gb1n(k-1)
  – Execution time factor g = b`1/b1

                                                  8
                  Stability
• Real system model, updated based on
  variation parameter
• G(z) = g / (z – (1 – g)
  – |1 – g| < 1
• Poles need to be within unit circle
• Stable as long as 0 < g < 2



                                        9
           Steady State Error
• The steady state of the system is derived
  – limz->1 (z - 1) G(z) Rref (z / (z - 1))
  – limz->1 gz / (z – (1-g)) Rref
  – Rref
• Thus the system is guaranteed to achieve
  the response time if the system is stable



                                              10
                Settling Time
• r(k) = (1 – g) r(k – 1) + gRref
• Settles when the systems converges to
  Rref ± 0.05
• The number of control periods required to
  settle is
  – k ≥ ln 0.05 / ln |1 - g|



                                              11
            Implementation
• Assumption that updates arrive in 2 – 5
  second intervals
  – Current work to relax this assumption
• 1 second set point




                                            12
                  Baselines
• OPEN
  – Fixed job budget
  – Can guarantee response time when estimated
    execution time is correct
  – May violate timing when execution time is
    underestimated
• Ad Hoc
  – Heuristic-based adaptive controller
  – Fixed step increments each control period based on
    whether response time is above or below set point.

                                                         13
             Control Accuracy
• Starts using design time
  execution estimates (ie
  g=1)
• At time 1000, execution
  time increases to g=1.4
• At time 2000, g=1.8
• OPEN fails to handle
  changes in execution
  time
• PI controller meets
  deadlines and settling
  time design

                                14
          Comparison to Ad Hoc




• Starts out with g = 0.6
• At 800s, g = 1
• Ad Hoc takes 380s to settle vs 100s for PI controller


                                                          15
     Quality of Service (QoS)
• Open only
  considered when
  g≤1
• PI controller
  offers better QoS
  than both OPEN
  and Ad Hoc


                                16
        Different Time Factors
• Relationship between
  response time and
  execution factor
• When g=2.6, the
  controller oscillates
• The response time
  stays close to the set
  point when the
  execution time factor
  is between 0 and 2

                                 17
           Settling Time
        Results vs Theoretical
• The experimental
  results are very
  close to the
  theoretical values
  predicted
• Experiments
  validate the
  theoretical analysis

                                 18
                Summary
• Real-time information dissemination is
  used to share information in timely manner
  – Valuable Information at the Right Time (VIRT)
• PI controller maintains response time
  guarantees within settling time constraints
  with no steady state error
• Superior performance to OPEN and Ad
  Hoc (heuristic) controllers
                                                19
           ECE 555
 Real-Time Embedded Systems

Chronos: Feedback Control of a
    Real Database System
         Performance
          Presented by
           Ben Taylor
                                 20
                  Outline
• Introduction
  – What does a real-time database offer that
    existing databases do not?
• Solutions
  – Feedback controller
  – Adaptive update policy
• Results and Performance
• Summary
                                                21
    Why real-time databases?
• Existing databases have no notion of data
  freshness or timing deadlines
  – Stock trading system needs to keep prices up
    to date while supporting reasonable response
    times
• Need soft real-time constraints on
  transactions while maintaining up-to-date
  data

                                               22
Architecture Overview




                        23
           Controller Design
• If the system is overloaded the queue will
  tend toward unbounded growth
• If the system is underused, the queue size
  will tend to be small or empty
• The controlled variable is the service delay
• The manipulated variable is the ready
  queue size
  – If queue is full, transactions are not accepted
                                                      24
 Feedback Controller Overview
• At kth sampling, calculate delay error
  e(k) = Ds – d(k)
• Compute δq(k) based on e(k)
• If δq(k) < 0
  – Adjust adaptive update policy by increasing
    the period of cold data and increase δq(k) by
    (p[i]new – p[i])/p[i] until δq(k) ≥ 0 or period max
• q(k) = q(k-1) + δq(k)
  – 0 ≤ q(k) ≤ max_qsize
                                                      25
          Freshness Adaptation
• Control of data di, period of p[i]
• Initially p[i] = 0.5 avi[i], absolute validity interval
• AUR[i] = Access Frequency[i] / Update
  Frequency[i]
   – di is hot if AUR[i] ≥ 1
   – Otherwise it is cold
• When increasing p[i]new = min(p[i]/AUR[i], Pmax)
• After each update period, fvi[i]new = 2p[i]new,
  where fvi[i] = avi[i] intially
• avi[i] ≤ fvi[i]new ≤ 2Pmax
                                                            26
       System Identification
• Used to model relationship between the
  service delay and the queue size
• PI controller in the z domain
• Root Locus method in Matlab, similar to
  EUCON, to show controller is stable




                                            27
              Performance
• Open - Pure Berkeley DB
  – Standard state-of-the-art database
• AC – Ad-hod Admission Control
  – Admission control in proportion to error
• FC-C – Feedback Control AC
  – Admission control with feedback loop
• FC-CU – Feedback Control AC + AUP
  – Adaptive temporal updates and admission
    control with feedback loop
                                               28
  Performance Comparison




Ds = 2s, Do = 2.5s, Dt = 100s, Pmax = 5s

                                           29
              Summary
• Real-time databases need to balance
  timely response with fresh data
• Designed feedback controller to manage
  backlog in system
• Adaptive update policy to manage
  freshness based on temporal data access
  and update patterns


                                        30
            Comparisons
• Both use system identification for
  controller design, different models
• Chronos system maintains data freshness,
  a component not in the Information
  dissemination system
• Chronos system controller handles
  concurrency issues not present in
  Information dissemination system
                                         31
                Critiques
• Both assume inter-arrival times of a limited
  window (2s, 5s) and (1s, 3s)
• Chronos systems states that workloads
  outside of operating range is reserved for
  a future work
• Information Dissemination assumes a
  given number of subscriptions will have
  the same cost as a different set of
  subscriptions the same size
                                             32
Questions




            33

				
DOCUMENT INFO