Alina Weffers-Albu, [e-mail]

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

• Terminal QoS
• Alina Weffers-Albu

2
Contents

• Context
• Progress
• Project definition Goals, Approach
• Characterization of CS sequences
• Stable State Theorem
• Execution streaming chains - dependency on input.

3
Context - QoS in IN-Home Digital Networks
Aim provide guaranteed and optimised Quality of
Service (QoS) for interconnected real-time
embedded systems.
Network QoS Reliability, Delay, Jitter, Bandwidth
.
Terminal QoS Performance
4

Context - Description of Analyzed Systems
5
Context - Description of Analyzed Systems
Components

• Data driven. Execution determined by
• Availability of necessary input
• Priority of component task
• Time driven. Execution determined by
• Availability of necessary input. (Or NOT)
• Priority
• Periodicity.

6
Context - Description of Analyzed Systems
Components

• Both types. Execution determined by
• Average computation time.
• n-gtm relation between input and output.
• If m variable average m or distribution over
  time for the values of m.
• Average times needed to get each input FP/EP.
• Average times needed to produce each output
  FP/EP.
• Average suspension time (if task with execution
  deferral due to cooperation with hardware).

7
Previous results

• Performed a literature survey on QoS work
• Studied ways of estimating the overhead
  introduced by CS during the execution of
  streaming chains.
• Provided a method for the calculation of the
  overhead introduced by CS.
• Method based on an observation regarding the
  execution of streaming chains. Method tested on
  single case study.

8
Progress

• Expanded approach previously tested on particular
  case to a more general context - tests on other
  types of components, different priorities
  assignment.
• Formulate Stable Phase Theorem, distinguished 7
  separate cases of interest for proof.
• gt Approach for control and optimization of
  performance parameters by formulating corollaries
  deduced from the proof.
• Proof for first case, lemmas, corollaries.
• Studied influence of input on the execution
  pattern of a streaming chain.
• Defined goals and approach for PhD project.(not
  restricted to CPU, but also memory, bus
  correlation of events sequences)

9
Goals

• Terminal QoS
• Performance
• Predictability of the system
• Goals
• Prediction of performance quality parameters for
  a given system.
•  
• Control performance quality parameters - find
  good practices of design for the system so that
  its resources needs can be satisfied on the
  physical platform.

10
Approach

• Study and model the dynamic behavior of a given
  system
• gt prediction control of performance quality
  parameters
• Behavior characterization in terms of the events
  that occur during the execution of the system.
• Events in our study
• Currently buffer handling operations, context
  switches,
• Future memory accesses (to be extended).

• Theoretical framework to model the sequences of
  events.
• Derive characteristics of the sequences of events
  gt meaningful abstractions.(Ex repetitive
  patterns, bounds)
• Identify conditions under which a sequence of
  events adopts a particular characteristic.
• Identify correlations and dependencies between
  sequences of events (CS, memory accesses, events
  related to bus utilization).

11
Performance Quality Parameters.
Buffer size
Packet size
Activation Times
Priority setting
Resource Utilization (RU) for CPU, memory, bus
feasibility check on the physical platform at
hand.
Activation Times (AT) provide modeling basis
for the sequence of context switches (CS).
Response Times (RT) prediction/control of
deadline misses.
Number of Context Switches (NCS) overhead
induced by the composed execution of components.
Required buffer space
12
Characterization of CS sequences.

• Hypothesis
• Let C1, C2, C3, , Cn be a chain of components
  communicating through a set of
• queues. The execution of the chain, after an
  initialization phase adopts a
• repetitive pattern of execution.
• Conditions under which the above statement holds
  in progress to be explored.
• Examples
• input - constant rate and sufficiently long,
  components designed such that their execution in
  the chain does not lead to deadlock.

13
Two case studies

• VO
• Time driven
• 1-gt2

• SSE
• -Data driven
• 1-gt1

• FRead
• Data driven with execution deferral

• VDec
• Data driven
• 1-gtm, m variable

NCS Stable Phase Calculated 900 NCS Stable
Phase Measured 895

• FRead
• Data driven with execution deferral

• C2
• -Data driven
• 2-gt3

• C1-C8
• -Data driven
• 1-gt1

NCS Stable Phase Calculated 245 NCS Stable
Phase Measured 245
P
Components
FRead
C1
C2
C3
C4
C5
C6
C7
14
Stable State Theorem.Cases of interest for proof.

• C1, C2, C3, , CN chain of components
  communicating through a set of queues (slide 4)
• N data-driven components (1-1).
• N data-driven components (n-m).
• C1 data-driven component with execution deferral
  (1-1), C2, C3, , CN data-driven components
  (n-m).
• C1, C2, C3, , CN-1 data-driven components (n-m),
  CN time-driven component (n-m).
• C1 time-driven component (n-m), C2, C3, , CN
  data-driven components (n-m).
• C1 time-driven component (n-m), C2, C3, , CN-1
  data-driven components (n-m), CN time-driven
  component (n-m).
• C1 data-driven component with execution deferral
  (1-1), C2, C3, , CN-1 data-driven components
  (n-m), CN time-driven component (n-m).

15
Stable State Theorem. 1-1 data-driven components.

• Let C1, C2, C3, , CN be a chain of data-driven
  components communicating
• through a set of queues (slide 4). The relation
  between input and output for all
• components is 1-1. The execution of the chain,
  after a finite initialization phase
• adopts a repetitive pattern of execution.
• Conditions input - constant rate and
  sufficiently long, components designed such that
  their execution in the chain does not lead to
  deadlock.
• Lemma 1 At stable state the execution of all
  components is dependent on the execution of the
  component with the lowest priority. (The
  component with the lowest priority in the chain
  is driving component).
• Lemma 2 If Cm is the driving component in the
  chain then ? i 1 i lt m, L(FQi) S(FQi) ? ?
  i m i lt N, L(FQi) 0.
• Corollary 1 The minimum buffer length necessary
  to ensure the repetitive execution is 1.
• Corollary 2 The NCS can be reduced by assigning
  priorities in a descending order from left to
  right.
• Corollary 3 The length of the initialization
  time can be reduced by reducing the buffers
  length.

16
Characterization of CS sequences.
Initialization phase C1 executes until output
FQ is filled gt C1 - Blocked (b).

• Chain
• - N data driven components
• - n-gtm 1-gt1
• - priorities in descending order.

C2(p)C1(b), C2(p)C1(b), , until C2(b) (FQ
filled, EQ empty)C1(b),
C3(p)C2(p)C1(b) C2(b), C3(p) C2(p)C1(b) C2(b)
C3(p) C2(p)C1(b) C2(b), C3(b)
CN(p)CN-1(p) C2(p)C1(b)C2(b)CN-1(b),
Stable phase CN(p)CN-1(p) C2(p)C1(b)C2(b)CN-1(b
),
17
Influence of input on the execution of a chain

• Correlation between pattern in MPEG input and
  pattern of execution.
• Characterization of input stream
• Guidelines for intelligently choosing the size of
  the packets in order to increase predictability
  for components with variable output.

18
Other activities

• Papers
• NCS Calculation Method for Streaming
  Applications. Proceedings of the 5th PROGRESS
  Symposium on Embedded Systems
• A Characterization of Streaming Applications
  Executions (submitted to the Design, Automation,
  and Test in Europe 2005 Conference)
• In process of writing paper with Radu Dobrin
  University of Malardalen Sweden
• Cooperations
• Malardalen University, Sweden - Gerhard Fohler,
  Radu Dobrin
• Carnegie Mellon, SEI Kurt Wallnau, Mark Klein
• Presenting my work
• Poster 5th PROGRESS Symposium on Embedded
  Systems, October 2004
• Presentations for SAN group(TU/e), OASIS cluster
  (Philips Research), Carnegie Mellon SEI, Gerhard
  Fohler (Malardalen University)

19
Other activities

• Presenting my work
• Liesbeth Steffens - Philips Research Laboratories
• Reinder Bril - Eindhoven University of Technology
• Clara Otero-Perez - Philips Research Laboratories
• Laurentiu Papalau - Philips Research Laboratories
• Giel van Doren - Philips Research Laboratories
• Dietwig Lowet - Philips Research Laboratories
• Sjir van Loo -  Philips Research Laboratories
• Jan van der Wal - Eindhoven University of
  Technology
• Clemens Wust -  Philips Research Laboratories
• Marco Bekooij -  Philips Research Laboratories
• Jeffrey Kang - Philips Research Laboratories
• Saianath Karlapalem - Singapore