a mixed model for estimating the probabilistic worst case execution time cristian maxim*, adriana...

24
A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt, France *Airbus, Toulouse Open problems in real-time computing April 4th, 2014, ULB, Brussels, Belgium

Upload: miranda-kibble

Post on 29-Mar-2015

217 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE

EXECUTION TIME

Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN

INRIA Paris-Rocquencourt, France

*Airbus, Toulouse

Open problems in real-time computing April 4th, 2014, ULB, Brussels, Belgium

Page 2: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Summary

• About probabilities

• Measurement-based probabilistic time analysis (MBPTA)

• Genetic algorithms

• Our mixed model

WHY MBPTA NEEDS to be IMPROVED?

Page 3: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Probabilities

• What is a distribution function?

• What is a probabilistic real time system?

• Central limit theorem

• Extreme value theory

• Independence and identical distribution (i.i.d.)

Page 4: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

What is a probability distribution function?

• A function that gives the probability of a random variable to be equal to a given value

• Continuos random variable Probability density function (pdf)

Probabilities

Page 5: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

What is a probability distribution function?

• A function that gives the probability of a random variable to be equal to a given value

• Discrete random variable Probability mass function (pmf)

Probabilities

𝒞𝑖=( 1 3 70.2 0.5 0.3)

Page 6: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Cumulative distribution function (cdf)• It describes the probability that a real-valued random

variable X with a given probability distribution will be found at a value less than or equal to x

Probabilities

Continuous random variable Discrete random variable

Page 7: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Probabilistic real-time systems (pRTS)

• pRTS – a real time system with at least one of the parameters represented as a random variable

• Model of real time system:

Probabilities

task (task set)

Offset

WCET

Period

Deadline

Page 8: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Probabilistic real-time systems (pRTS)

• One parameter described by a random variable:

• • Most known• Studied by Diaz, Cucu and others.

• • Practical example: two cars backing up

• •

Probabilities

Page 9: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Probabilistic real-time systems (pRTS)

• Example:

Probabilities

Page 10: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Central Limit Theorem (CLT)• Lehoczky [1992, 1995], Tia [1995], Broster [2002]

• It states that the sample mean is aproximatively a Gaussian distribution, given a sufficiently large sample. (gaussian distribution = normal distribution)

Probabilities

Tail

Page 11: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Extreme value theory (EVT)• Estimates the probability of occurrence of extreme events, when their

distribution function is unknown, based on sequences of observations. • If the distribution of rescaled maxima converges, then the limit G(x) is one

of the three following types:

Probabilities

Gumbel pdf

Page 12: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Independence and identical distribution (i.i.d.)

• In order to use EVT or CLT, the input data for these techniques has to be:

• Independent

• Identical distributed

Probabilities

Page 13: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Probabilistic Worst Case Execution Time (pWCET)

• The pWCET is an upper bound on the execution times of all possible jobs of the task

Probabilities

Page 14: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Measurement-based probabilistic timing analysis (MBPTA)• Steps of applying EVT (single-path programs)

Observations

Grouping

Fitting

Comparison

Tail extension

MBPTA

- Tested to be i.i.d.

- A fair amount of observation is needed

- The input data should vary

Page 15: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Measurement-based probabilistic timing analysis (MBPTA)• Steps of applying EVT (single-path programs)

Observations

Grouping

Fitting

Comparison

Tail extension

MBPTA

Block maxima technique

Page 16: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Measurement-based probabilistic timing analysis (MBPTA)• Steps of applying EVT (single-path programs)

Observations

Grouping

Fitting

Comparison

Tail extension

MBPTA

Finding the parameters for the Gumble distribution

• Location - μ

• Scale - β

• Shape -α

Page 17: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Measurement-based probabilistic timing analysis (MBPTA)• Steps of applying EVT (single-path programs)

Observations

Grouping

Fitting

Comparison

Tail extension

MBPTA

Page 18: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Measurement-based probabilistic timing analysis (MBPTA)• Steps of applying EVT (single-path programs)

Observations

Grouping

Fitting

Comparison

Tail extension

MBPTA

Page 19: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Measurement-based probabilistic timing analysis (MBPTA)

• The MBPTA ensures safeness (tight and pessimistic bound on WCET) with respect to the input data

How we build representative input data with respect to the WCET?

MBPTA

Page 20: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Genetic Algorithms

Genetic Algorithms

• Belong to the larger class of evolutionary algorithms

• Used in optimization problems in order to get better solutions

• In our case – we use it to get a large and diversified number of inputs in order to access all paths of a program

Page 21: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Genetic Algorithms

Genetic Algorithms

Page 22: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

A mixed model for estimating the probabilistic worst case execution time

Page 23: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Conclusion

• Experiments needed

• Verification of i.i.d. for both inputs and execution times

• Is there any corelation between the inputs and the execution times?

Page 24: A MIXED MODEL FOR ESTIMATING THE PROBABILISTIC WORST CASE EXECUTION TIME Cristian MAXIM*, Adriana GOGONEL, Liliana CUCU-GROSJEAN INRIA Paris-Rocquencourt,

Thank you for your attention