Alina Albu, m'a'albutue'nl

1
Quality of Service for In-Home Digital Networks
PROGRESS PROJECT EES.5653

• Terminal QoS
• M.A. Albu

2
Contents

• Summary work
• Terminal QoS
• Collaboration with MRM project
• Number of context switches estimation method
• Future work

3
Summary work

• Literature survey on QoS work
• QoS overview and classification of QoS
  improvement techniques internal report
• Number of context switches estimation method
• 1st approach statistical approach
• 2nd approach min-max method
• 3rd approach average method

4
Terminal QoS

• QoS determined by resource management of the
  system in discussion
• Terminal resources under investigation
• CPU,
• Memory,
• Bus

5
Collaboration with MRM project

• Why MRM?
• QoS related closely to resource management.
• MRM is concerned with resources management
  aspects in the context of a terminal.
• MRM provides opportunities for inspiration,
  validation of my work
• Aim MRM
• provide methods and means for an integrated
  approach to resource management in multi-resource
  systems.
• The integrated approach has to meet at least the
  following requirements
• The resource management infrastructure should be
  able to provide resource guarantees to the
  building blocks of application functionality.
• Individual building blocks should be able to
  limit or prevent resource insufficiencies, by
  dealing with insufficient resources in a graceful
  and predictable way.

6
Collaboration with MRM project Rationale

• This leads to the need for developing methods for
  estimating the necessary of resources for the
  building blocks of application but also for
  predicting resources necessary for the composed
  execution of these blocks.
• Why performance composition? - Just adding clock
  cycles of the involved components wont do.
• Method for the estimation of the number of
  context switches occurring during the execution
  of a streaming application.
• Current experimentation setting
• HW Trimedia (TM 1300)
• incorporates a media processor for
  high-performance multimedia applications that
  deal with high-quality video and audio.
• SW TSSA

7
MRM project - TSSA

• TSSA TriMedia Streaming Software Architecture

8
Reasons for context switch occurrence

• Blocking. The execution of a task blocks because
  of the following reasons
• o        Communication with the PC host (ex
  FRead)
• o        Unfavorable status of the queues
• -         input full packets queue (IFPQ) is
  empty (no input)
• -        output full packets queue (OFPQ) is
  full (task cannot output packets for the
  moment)
• -         output empty packets queue (OEPQ) is
  empty (task cannot output packets for the
  moment)
• Preemption. The execution of a task is preempted
  by another task with a higher priority.
• Task execution end. The execution of a task with
  high priority has ended (no preemption or
  blocking) and the resources are allocated to
  another task.

9
NCS Estimation MethodProperties of streaming
applications executions

• Property1 Running streaming applications, after
  an initialization phase, adopt a pattern of
  execution that repeats after a specific interval
  of time (hyperperiod). The repetitive execution
  is caused by the differences in the components
  rates of production/consumption of full/empty
  packets.
• Execution consists of 3 phases initialization,
  stable-state, finalization
• gt by knowing the NCS occuring during
  initialization, finalization and during a
  hyperperiod of the steady-state, we obtain the
  total NCS
• Property2 When one of the components in the
  streaming chain is periodic, when other
  components depend on it in execution, their tasks
  will execute periodic.

10
NCS Estimation Method Case study description
11
NCS Estimation MethodSteps 1 4
initialization, finalization phases

• Step 1.
• Initialization phase
• End phase 169.261 ms
• Duration 169.261 ms
•  Step 2.
• Initialization phase
• NCS_initializationPhase(FRead) 26
• NCS_initializationPhase (VDec) 22
• NCS_initializationPhase (VRendVO) 6
• Step 3.
• Finalization phase
• Beginning phase 4539.029 ms
• Duration 1569.428 ms
• Step 4.
• Finalization phase
• NCS_finalizationPhase(FRead) 8
• NCS_finalizationPhase(VDec) 356
• NCS_finalizationPhase(VRendVO) 94

12
NCS Estimation Method Steps 5 6 stable-state
phase, production/consumption rates

• Step 5.
•   Stable state
• Beginning phase 169.261 ms
• End phase 4539.029 ms
• Duration4369.768 ms
• Step 6.
• Identify for each component the full packets
  production rate (FPPR), the full packets
  consumption rate (FPCR), and the empty packets
  production rate (EPPR).
• Priority FPPR FPCR EPPR T
  AT CT CEPT (ms)
• FRead 90 2.2
  
  2.524
• Vdec 70 4.6 17.9
  
  4.5 2
• VrendVO 80 16.3
  16.3
  0.056 32.6
• - measurements of components rates and
  computation times in isolation.

13
NCS Estimation Method Steps 5 6 stable-state
phase, production/consumption rates

• Step 5.
•   Stable state
• Beginning phase 169.261 ms
• End phase 4539.029 ms
• Duration4369.768 ms
• Step 6.
• Identify for each component the full packets
  production rate (FPPR), the full packets
  consumption rate (FPCR), and the empty packets
  production rate (EPPR).
• Priority FPPR FPCR EPPR
  T AT CT CEPT (ms)
• FRead 90 2.2
  
  2.524
• Vdec 70 4.6 17.9
  
  4.5 2
• VrendVO 80 16.3
  2FPPR(VO) 16.3
  0.056 32.6

14
NCS Estimation Method Steps 5 6 stable-state
phase, production/consumption rates

• Step 5.
•   Stable state
• Beginning phase 169.261 ms
• End phase 4539.029 ms
• Duration4369.768 ms
• Step 6.
• Identify for each component the full packets
  production rate (FPPR), the full packets
  consumption rate (FPCR), and the empty packets
  production rate (EPPR).
• Priority FPPR FPCR EPPR
  T AT CT CEPT (ms)
• FRead 90 2.2
  
  2.524
• Vdec 70 4.6 17.9
  
  4.5 2
• VrendVO 80 16.3
  2FPPR(VO) 2FPPR(VO) 16.3
  0.056 32.6

15
NCS Estimation Method Steps 5 6 stable-state
phase, production/consumption rates

• Step 5.
•   Stable state
• Beginning phase 169.261 ms
• End phase 4539.029 ms
• Duration4369.768 ms
• Step 6.
• Identify for each component the full packets
  production rate (FPPR), the full packets
  consumption rate (FPCR), and the empty packets
  production rate (EPPR).
• Priority FPPR FPCR EPPR
  T AT CT
  CEPT(ms)
• FRead 90 2.2
  
  2.524
• Vdec 70 4.6 17.9
  4FPPR(VDec)
  4.5 2
• VrendVO 80 16.3
  2FPPR(VO) 2FPPR(VO) 16.3
  0.056 32.6

16
NCS Estimation Method Steps 5 6 stable-state
phase, production/consumption rates

• Step 5.
•   Stable state
• Beginning phase 169.261 ms
• End phase 4539.029 ms
• Duration4369.768 ms
• Step 6.
• Identify for each component the full packets
  production rate (FPPR), the full packets
  consumption rate (FPCR), and the empty packets
  production rate (EPPR).
• Priority FPPR FPCR EPPR
  T AT CT
  CEPT(ms)
• FRead 90 2.2 -
  -
  2.524 -
• Vdec 70 4.6 17.9
  4FPPR(VDec)
  4.5 2
• VrendVO 80 16.3
  2FPPR(VO) 2FPPR(VO) 16.3
  0.056 32.6

17
NCS Estimation Method Step 7 dependencies
between components

• Step 7
• a - Determine the dependencies in the execution
  of the components by taking into consideration
  FPPR, FPCR and EPPR for each component.
• b - Determining the dependencies in the
  execution of the components, leads to determining
  the period (T(Ti)) of each task Ti on which the
  components Ci are mapped.
• VDec
• a - FPPR (VDec) gt FPCR (VRendVO) (gt rate
  higher) gt OFPQ (VDec) at stable state is full gt
  OEPQ (VDec) is empty gt VDec depends on VRendVO
  to produce 1 EP so that VDec can produce 1 FP gt
•  
• FPPR (VDec) EPPR (VRendVO) 2
  FPPR(VrendVO) 2 T(VrendVO) 32.6 ms.
• gt
•  
• b - Since T(Vdec) FPPR(VDec) gt T(Vdec) 32.6
  ms

18
NCS Estimation Method Step 7 dependencies
between components

• Step 7
• a - Determine the dependencies in the execution
  of the components by taking into consideration
  FPPR, FPCR and EPPR for each component.
• b - Determining the dependencies in the
  execution of the components, leads to determining
  the period (T(Ti)) of each task Ti on which the
  components Ci are mapped.
• FRead
• a FPPR (FRead) gt FPCR (VDec) (gt rate higher)
  gt OFPQ (FRead) at stable state is full gt OEPQ
  (FRead) is empty gt FRead depends on VDec to
  produce 1 EP so that FRead can produce 1 FP gt
•  
• FPPR (FRead) EPPR (VDec) 4
  FPPR(VDec) 4 T(VDec) 42T(VrendVO)
  130.4 ms
• gt
•  
• b - Since T(FRead) FPPR(FRead) gt T(FRead)
  130.4 ms

19
NCS Estimation Method Step 7 dependencies
between components

• Step 7
• Priority FPPR FPCR EPPR
  T AT CT
  CEPT(ms)
• FRead 90 2.2 -
  -
  2.524 -
• Vdec 70 4.6 17.9
  4FPPR(VDec)
  4.5 2
• VrendVO 80 16.3
  2FPPR(VO) 2FPPR(VO) 16.3
  0.056 32.6

20
NCS Estimation Method Step 7 dependencies
between components

• Step 7
• Priority FPPR FPCR EPPR T
  AT CT CEPT (ms)
• FRead 90 2.2 -
  -
  2.524 -
• Vdec 70 4.6 17.9
  130.4
  4.5 2
• VrendVO 80 16.3 32.6
  32.6 16.3
  0.056 32.6

21
NCS Estimation Method Step 7 dependencies
between components

• Step 7
• Priority FPPR FPCR EPPR T
  AT CT CEPT (ms)
• FRead 90 130.4 -
  -
  2.524 -
• Vdec 70 32.6 17.9
  130.4
  4.5 2
• VrendVO 80 16.3 32.6
  32.6 16.3
  0.056 32.6

22
NCS Estimation Method Step 7 dependencies
between components

• Step 7
• Priority FPPR FPCR EPPR T
  AT CT CEPT (ms)
• FRead 90 130.4 -
  - 130.4
  2.524 -
• Vdec 70 32.6 17.9
  130.4 32.6
  4.5 2
• VrendVO 80 16.3 32.6
  32.6 16.3
  0.056 32.6

23
NCS Estimation Method Step 8 hyperperiod
length, number of hyperperiods

• Step 8.
• Identify hyperperiod length.
• CIS (component index set) the set of natural
  numbers that serve as indexes for components
• in a streaming chain. The indexes of components
  will be equal with the indexes of the tasks on
• which the components are mapped at execution.
• HL max T(Ti) 8 T(VRendVO) 130.4 ms
• i?CIS
•  
• Duration stable phase 4369.768 ms
• gt average number of hyperperiods during stable
  phase
•  
• HN ?Duration_stableStatePhase/HL?
  ?4369.768/HL? ?4369.768/130.4? 34

24
NCS Estimation Method Step 9 NCS due to
blocking

• Step 9.
•  
• Determine the NCS due to blocking.
•  
• FRead
•  
• FRead blocks 4 times for each packet that it
  delivers due to communication with the PC host
  and has its period equal with the hyperperiod
  (because it only gets to deliveres 1 full packet
  during the hyperperiod after which it blocks
  until the next hyperperiod) gt
•  
• NCS_blocking(FRead) ?HL / T(Ti)?
  NCS_inherentBlocking(FRead) 1 4 5

25
NCS Estimation Method Step 9 NCS due to
blocking

• Step 9.
•  
• VDec
• VDec delivers 1 full packet every time it
  is activated after which it is blocked.
• VDec is activated periodically and that
  its period fits 4 time during the hyperperiod gt
• VDec is activated 4 times during the
  hyperperiod gt VDec is blocked 4 times during
  the hyperperiod.
•  
• NCS_blocking(VDec) ?HL / T(Ti)?
  NCS_inherentBlocking(VDec) ?HL / T(VDEc)? 0
• 
  ?8T(VRendVO) / 2T(VRendVO)?
• 4
•  
• VRendVO
• VRendVO is the component that does not depend on
  any other component in its execution, and has no
  inherent blockings gt it does not block.
•  
• NCS_blocking(VRendVO) 0

26
NCS Estimation Method Step 10 NCS due to
normal execution end.

• Step 10.
•  
• Determine the NCS due to normal execution.
•  
• Applies only to components that are not preempted
  and do not depend on any other component in its
  execution, thus do not get blocked.
•  
• gt Applies only to VRendVO
•  
• NCS_normalExecutionEnd(VRendVO) ?HL /
  T(VRendVO)?
• 
  ?8T(VRendVO) / T(VRendVO)?
• 
  8
•  

27
NCS Estimation Method Step 11 Task activation
time (AT)

• Step 11.
• For each component by considering the
  dependencies dictated by the rates of
  production/consumption packets, calculate the
  first AT in the hyperperiod.
• In general  For each component Ci mapped on task
  Ti
•  1. If ? j ? CIS Ci dependent on Cj
• If P(Ti) lt P(Tj) AT(Ti) AT(Tj) (N-1)
  FPPR(Tj) CT(Tj) if Ti depends on Tj to
  release N FP.
• (N-1) EPPR(Tj) CT(Tj) if Ti depends on
  Tj to release N EP.

N4
CT
28
NCS Estimation Method Step 11 Task activation
time (AT)

• If P(Ti) gt P(Tj) AT(Ti) AT(Tj) CFPT(Tj)
  (N-1) FPPR(Tj) if Ti depends on Tj to release
  N FP.
  
  
  
  
  
  
• CEPT(Tj) (N-1) EPPR(Tj)
  if Ti depends on Tj to release
  N EP.

29
NCS Estimation Method Step 11 Task activation
time (AT)

• For the current case study
• FRead FRead dependent on Vdec to release 1 EP
  gt
•  
• P(FRead) gt P(VDec) gt AT(FRead) AT(VDec)
  (1-1)EPPR(Vdec) CEPT (VDec)
• 
  AT(VDec) CEPT(VDec)
• 
  AT(VDec) 2ms.
• Vdec Vdec dependent on VRendVO to release 1 EP
  gt
•  
• P(VRendVO) gt P(VDec) gt
• AT(VDec) AT(VRendVO) (1-1)EPPR(VRendVO) CT
  (VRendVO)
• AT(VRendVO) CT (VRendVO)
• AT(VRendVO) 0.056 ms.
• 0.056 ms
•  
• First AT(VRendVO) 0 relative to the beginning
  of the hyperperiod.
•  
• gt AT(FRead) 0.056 ms 2 ms 2.056 ms

30
NCS Estimation Method Step 11 Task activation
time (AT)

• Step 11
• Priority FPPR FPCR EPPR T
  AT CT CEPT (ms)
• FRead 90 130.4 -
  - 130.4
  2.524 -
• Vdec 70 32.6 17.9
  130.4 32.6
  4.5 2
• VrendVO 80 16.3 32.6
  32.6 16.3
  0.056 32.6

31
NCS Estimation Method Step 11 Task activation
time (AT)

• Step 11
• Priority FPPR FPCR EPPR T
  AT CT CEPT (ms)
• FRead 90 130.4 -
  - 130.4 2.056 2.524
  -
• Vdec 70 32.6 17.9
  130.4 32.6 0.056 4.5
  2
• VrendVO 80 16.3 32.6
  32.6 16.3 0
  0.056 32.6

32
NCS Estimation Method Task response time, NCS
due to preemptions

• P(Ti) gt P(Tj) ? AT(Ti) ? (AT(Tj), AT(Tj)
  CT(Tj)) gt Ti preempts Tj.

AT(Tj)
CT(Tj)
R0(Tj) CT(Tj)
33
NCS Estimation Method Task response time, NCS
due to preemptions

• P(Ti) gt P(Tj) ? AT(Ti) ? (AT(Tj), AT(Tj)
  CT(Tj)) gt Ti preempts Tj.

CT(Ti)
AT(Tj)
AT(Ti)
R1(Tj) Ro(Tj)CT(Ti)
34
NCS Estimation Method Task response time, NCS
due to preemptions

• P(Ti) gt P(Tj) ? AT(Ti) ? (AT(Tj), AT(Tj)
  CT(Tj)) gt Ti preempts Tj.

T(Ti)
CT(Ti)
AT(Tj)
AT(Ti)
AT(Ti)
R1(Tj) Ro(Tj)2CT(Ti)
35
NCS Estimation Method Task response time, NCS
due to preemptions

• P(Ti) gt P(Tj) ? AT(Ti) ? (AT(Tj), AT(Tj)
  CT(Tj)) gt Ti preempts Tj.
• NCS_preemption(Tj) ?Ro (Tj)/T(Ti)?
• R1(Tj) NCS_preemption(Tj)CT(Ti)

T(Ti)
CT(Ti)
AT(Tj)
AT(Ti)
AT(Ti)
R1(Tj) Ro(Tj)3CT(Ti)
AT(Ti)
36
NCS Estimation Method Step 12Task response
time, NCS due to preemptions

• Step 12
•  Calculate NCS_preemption for all components
• In general
• Rn(Ti) Rn-1 (Ti) ? ?Rn-1
  (Ti)/T(Tj)? CT(Tj),
  j ? k? CIS P(Tk) gt P(Ti) ? AT(Tk)
  ? (AT(Ti), AT(Ti) CT(Ti))
•  
•  where Ro initial response time, Ro(Ti)
  CT(Ti)   
• From here, the total number of context switches
  due to preemptions will be
•   
•  
• NCS_preemption(Ti) ? ?Rn-1
  (Ti)/Tj?
• j ? k? CIS
  P(Tk) gt P(Ti) ? AT(Tk) ? (AT(Ti), AT(Ti)
  CT(Ti))
•  

37
NCS Estimation Method Step 12Task response
time, NCS due to preemptions

• Step 12  For the current case study
• FRead FRead has he highest priority assigned gt
  never preempted.
• gt NCS(FRead)_preemption 0
• VDec
• P(VRendVO) gt P(VDec)
• AT(VRendVO), AT(VRendVO) ? (AT(VDec), AT(VDec)
  CT(VDec))
• gt VrendVO does not preempt VDec.
• P(FRead) gt P(VDec)
• AT(FRead), AT(FRead) ? (AT(VDec), AT(VDec)
  CT(VDec)) gt FRead preempts VDec
• NCS_preemption (VDec) 5
• VRendVO
• P(FRead) gt P(VRendVO)
• AT(FRead), AT(FRead) ? (AT(VRendVO), AT(VRendVO)
  CT(VRendVO))
• gt FRead does not preempt VRendVO.
• gt NCS_preemption(VRendVO)0

38
NCS Estimation Method Step 12Task response
time, NCS due to preemptions

• Step 12
•  For the current case study
• FRead FRead has he highest priority assigned gt
  never preempted.
• gt NCS(FRead)_preemption 0
• FRead - hybrid task that due to dependencies in
  execution activated every 130.4 ms. After
  activation it runs blocks 4 times /packet due to
  the communication with the PC host.
• The activation and blocking times after the first
  activation are the following
•  
• AT(FRead)1 CT1 0.121 ms , BT1 0.577 ms
• AT(FRead)2 CT2 0.152 ms, BT2 0.499 ms
•  AT(FRead)3 CT3 0.081 ms, BT3 0.126 ms
•  AT(FRead)4 CT4 0.09 ms, BT4 0.774 ms
•  AT(FRead)5 CT5 0.104 ms,
•  
• ð CT(FRead) ?CTi ?BTi 2.524 ms

39
NCS Estimation Method Step 12Task response
time, NCS due to preemptions

• VDec
• P(VRendVO) gt P(VDec)
• AT(VRendVO), AT(VRendVO) ? (AT(VDec), AT(VDec)
  CT(VDec))
• gt VrendVO does not preempt VDec.
• FRead - hybrid task gt each of its 5
  internal/inherent activations per period due to
  the
• communication with the PC host will be treated as
  independent tasks FRead1, FRead2, ,
• FRead5. The 5 tasks have the same period
  130.4 ms 8T(VRendVO) and have equal
• priorities with the priority assigned to FRead.
• P(FRead) gt P(VDec)
• ? AT(FRead), AT(FRead) ? (AT(VDec), AT(VDec)
  CT(VDec)) gt FRead preempts VDec

40
NCS Estimation Method Step 12Task response
time, NCS due to preemptions

•  Ro CT(VDec), Ro initial response time of
  VDec
• NCS_preemption(VDec)1 ?Ro
  (VDec)/T(FRead1)? ?4.5 / 130.4? 1
• R1 Ro ?Ro (VDec)/T(FRead1) ?
  CT(FRead1) 4.5 1 0.121 4.621 ms
•  
• NCS_preemption(VDec)2 ?R1
  (VDec)/T(FRead2)? ?4.621 / 130.4? 1
• R2 R1 ?R1 (VDec)/T(FRead2) ? CT(FRead2)
  4.773
•  
• NCS_preemption(VDec)3 ?R2
  (VDec)/T(FRead3)? ?4.773 / 130.4? 1 
• R3 R2 ?R2 (VDec)/T(FRead3) ? CT(FRead3)
  4.854 ms
•  
• NCS_preemption(VDec)4 ?R3
  (VDec)/T(FRead4)? ?4.854 / 130.4? 1
• R4 R3 ?R3 (VDec)/T(FRead4) ? CT(FRead4)
  4.944 ms
•   NCS_preemption(VDec)5 ?R4
  (VDec)/T(FRead5)? ?4.944 / 130.4? 1
• ð      the 5 tasks that compose the FRead hybrid
  task preempt VDec each 1 time gt
• ð      NCS_preemption (VDec) 5

41
NCS Estimation Method Step 12Task response
time, NCS due to preemptions

• VRendVO
• P(FRead) gt P(VRendVO)
• AT(FRead), AT(FRead) ? (AT(VRendVO), AT(VRendVO)
  CT(VRendVO))
• gt FRead does not preempt VRendVO.
• gt NCS_preemption(VRendVO)0

42
NCS Estimation Method Step 13 Total
NCS/hyperperiod

• Step 13
•  Determine NCS_total for each of the components
  involved
•  
• For each hyperperiod
•  NCS_total(Ci) NCS_blocking(Ci)
  NCS_preemption(Ci) NCS_normalExecutionEnd(Ci)
• gt Total NCS/hyperperiod
•  
• NCS_hyperperiod (FRead)  NCS_blocking(FRead)
  NCS_preemption(FRead) 5 05
• NCS_hyperperiod (VDec) NCS_blocking (VDec)
  NCS_preemption (VDec) 5 4 9
• NCS_hyperperiod (VRendVO) NCS_blocking VRendVO)
  NCS_preemption (VRendVO)
  NCS_normalExecutionEnd(VRe
  ndVO) 0 0 8 8
•   

43
NCS Estimation Method Step 14 Total NCS

• Step 14
•  
• Determine the total NCS during the entire
  execution of the streaming application.
•  
• We know that the average number of hyperperiods
  during stable phase HN 34 (from step 8)
•  
• gt Total estimated NCS
•  
• NCS_total(FRead) 534 26 8 204
  vs measured 207
• NCS_total (VDec) 934 22 356 684
  vs measured 679
• NCS_total (VRendVO) 834 6 94 372
  vs measured 362
•  
• Note Differences come from the fact that we work
  with averages in the components models which
  determines an average length for the hyperperiod
  and an average number of hyperperiods.

44
Future work

• Test method on more complex, realistic case
  studies
• Write paper describing the aforementioned
  findings
• Extend estimation method for applications
  containing multiple dependent/independent chains.
• Continue studies to finding ways to estimate the
  necessary of memory and bus for streaming
  applications.
• Continue studies on estimating necessary of
  resources streaming applications running on
  multiple processors platforms.