[PPT] - Submodel Pattern Extraction for Simulink Models James R. Cordy PowerPoint Presentation

SLIDE 1

Submodel Pattern Extraction for Simulink Models

James R. Cordy Queen’s University NECSIS Automotive Partnership Canada

SLIDE 2

MPE

near-miss clones

near-miss clone detection

Simone

analysis of GM models

pattern extraction

pattern evolution

SLIDE 3

MPE

near-miss clones

near-miss clone detection

Simone

analysis of GM models

pattern extraction

pattern evolution

SLIDE 4

Model Pattern Engineering discover, catalogue and formalize

submodel patterns

emergent

domain-specific, client-specific

SLIDE 5

discovery

analysis, identification methodology, techniques

classification

characterization, formalization notation, tooling, catalogues

application

deployment, analysis

rganization, documentation, use cases

SLIDE 6

why?

reuse in model development

standards/consistency analysis/enforcement

failure/change propagation in model maintenance

verification/test optimization

deployment variation/optimization

model product lines

SLIDE 7

MPE

near-miss clones

near-miss clone detection

Simone

analysis of GM models

pattern extraction

pattern evolution

SLIDE 8

code clones

copy-paste programming

efficient, widely used

problematic

SLIDE 9

code clones

type 1 - exact

bool ConfNextToken (char **p) { // skip white space while (1) switch (**p) { case '\t' : // ignore case ' ' : (*p)++; break; case '\0' : return FALSE; default : return TRUE; }; } bool ConfNextToken (char **p) { // skip white space while (1) switch (**p) { case '\t' : // ignore case ' ' : (*p)++; break; case '\0' : return FALSE; default : return TRUE; }; } [Roy, Cordy, Koschke SCP 2009]

SLIDE 10

code clones

type 1 - exact

bool ConfNextToken (char **p) { // skip white space while (1) switch (**p) { case '\t' : // ignore case ' ' : (*p)++; break; case '\0' : return FALSE; default : return TRUE; }; } bool ConfNextToken (char **p) { while (1) switch (**p) { case '\t': case ' ': // just skip (*p)++; break; case '\0': // eof return FALSE; default: // something we want return TRUE; }; } [Roy, Cordy, Koschke SCP 2009]

SLIDE 11

code clones

type 2 - renamed

bool ConfNextToken (char **p) { // skip white space while (1) switch (**p) { case '\t' : // ignore case ' ' : (*p)++; break; case '\0' : return FALSE; default : return TRUE; }; } bool NextToken (char **bp) { while (1) // not really switch (**bp) { case '\t': case ' ': // next (*bp)++; break; case '\0': return 0; default: return 1; }; } [Roy, Cordy, Koschke SCP 2009]

SLIDE 12

bool ConfNextToken (char **p) { // skip white space while (1) switch (**p) { case '\t' : // ignore case ' ' : (*p)++; break; case '\0' : return FALSE; default : return TRUE; }; }

code clones

type 3 - near miss

bool NextToken (char **bp) { while (1) // not really switch (**bp) { case '\t': (*bp)++ case ' ': break; case '\0': return 0; default: return 1; } } [Roy, Cordy, Koschke SCP 2009]

SLIDE 13

model clones

type 1 - exact

Tfmaxs 2 Tfmaxk 1 Torque Conversion 2/3*R*muk Ratio of static to kinetic mus/muk Fn 1 Tfmaxs 2 Tfmaxk 1 Torque Conversion 2/3*R*muk Ratio of static to kinetic mus/muk Fn 1

[Alalfi, Cordy, Dean, Stephan, Stevenson ICSM 2012] [Störrle SSM 2013]

SLIDE 14

model clones

type 2 - renamed

[Alalfi, Cordy, Dean, Stephan, Stevenson ICSM 2012] [Störrle SSM 2013]

SLIDE 15

model clones

type 3 - near miss

[Alalfi, Cordy, Dean, Stephan, Stevenson ICSM 2012] [Störrle SSM 2013]

SLIDE 16

MPE

near-miss clones

near-miss clone detection

Simone

analysis of GM models

pattern extraction

pattern evolution

SLIDE 17

ConQAT

graph-based model clone detection

[Deissenboeck et al. IWSC 2010]

SLIDE 18

graph flattening ignores hierarchical structure

problems with near-miss

SLIDE 19

graph flattening ignores hierarchical structure

problems with near-miss

SLIDE 20

code-based near-miss works well

NiCad, iClones, others

mature, accurate, efficient

handles unexpected differences

threshold-based, tunable

scalable

SLIDE 21

NiCad

parse - extract - normalize - diff threshold

Pretty-printed Potential Clones

1 2 3 4

Parsing & Potential Clone Extraction Original Code Base

1. Parse / Extract
2. Rename / Filter / Normalize

Renaming, Filtering, Normalization Normalized Potential Clones

1 2 3 4

Clone Classes

5.pc 23.pc 67.pc . . . 12.pc 17.pc 22.pc . . . 15.pc 18.pc 78.pc . . . 21.pc 63.pc 97.pc . . . 37.pc 39.pc 44.pc . . .

Choose Next Potential Clone as Exemplar

3. Clone Analysis

Comparable Size Potential Clone Cluster Pairwise Comparison with Exemplar (Repeat) Cluster Comparable Size PCs Normalized Potential Clones

1 2 3 4

[Roy, Cordy ICPC 2008]

SLIDE 22

crazy idea:

can we use near-miss text code methods

n graphical models?

“Models are source code too”

Mark Harman, keynote at SCAM 2010

[Harman SCAM 2010]

SLIDE 23

MPE

near-miss clones

near-miss clone detection

Simone

analysis of GM models

pattern extraction

pattern evolution

SLIDE 24

Simone

Simulink near-miss clone detection

experiment adapt NiCad near-miss code clone

detector to graphical models

validate vs. ConQAT

for types 1 & 2

hand validate type 3 (near-miss)

[Alalfi, Cordy,Dean, Stephan, Stevenson ICSM 2012]

SLIDE 25

Simulink

hybrid hardware/software models

widespread in industry

automotive, aerospace, embedded systems

mature and interesting at GM

SLIDE 26

Simulink

hierarchical models

[Alalfi, Cordy,Dean, Stephan, Stevenson ICSM 2012]

SLIDE 27

Challenge #1

code methods require text

NiCad requires a parser

Solution: grammar inference

n Simulink’s

internal form

... System { Name "onoff" Location [168, 385, 668, 686] Open on ModelBrowserVisibility off ModelBrowserWidth 200 ScreenColor "automatic" PaperOrientationi "landscape" PaperPositionMode "auto" PaperType "usletter" PaperUnits "inches" ZoomFactor "100" AutoZoom on ReportName "simulink-default.rpt" Block { BlockType DiscretePulseGenerator Name "Discrete Pulse\nGenerator" Position [45, 25, 75, 55] Amplitude "1" Period "2" PulseWidth "1" PhaseDelay "0" SampleTime "1" } Block { BlockType Product Name "Product" Ports [2, 1, 0, 0, 0] Position [145, 67, 175, 98] Inputs "2" SaturateOnIntegerOverflow on } ... } ...

SLIDE 28

Challenge #2

what granularity?

NiCad requires candidates for comparison

Simulink: model (too big) block (too small) system (just right!)

... System { Name "onoff" Location [168, 385, 668, 686] Open on ModelBrowserVisibility off ModelBrowserWidth 200 ScreenColor "automatic" PaperOrientationi "landscape" PaperPositionMode "auto" PaperType "usletter" PaperUnits "inches" ZoomFactor "100" AutoZoom on ReportName "simulink-default.rpt" Block { BlockType DiscretePulseGenerator Name "Discrete Pulse\nGenerator" Position [45, 25, 75, 55] Amplitude "1" Period "2" PulseWidth "1" PhaseDelay "0" SampleTime "1" } Block { BlockType Product Name "Product" Ports [2, 1, 0, 0, 0] Position [145, 67, 175, 98] Inputs "2" SaturateOnIntegerOverflow on } ... } ...

SLIDE 29

even with raw text, find some subsystem clones

but:

90% irrelevant Simulink internal “formatting” systems

some identical systems only 70% same

entirely missed exact copies displayed differently

neutral_up_down 1 mutually_exclusive

neutral up down validated_neutral validated_up validated_down

check_up

action reset checked_action

check_down

action reset checked_action

Goto1 [reset] From2 [reset] From1 [reset] reset 4 down 3 up 2 neutral 1

neutral_up_down

1 mutually_exclusive

neutral up down validated_neutral validated_up validated_down

check_up

action reset checked_action

check_down

action reset checked_action

Goto1 [reset] From2 [reset] From1 [reset] reset 4 down 3 up 2 neutral 1

SLIDE 30

Challenge #3

problems with “noise”

solution: “agile parsing” to

filter out irrelevant elements

... System { Name "onoff" Location [168, 385, 668, 686] Open on ModelBrowserVisibility off ModelBrowserWidth 200 ScreenColor "automatic" PaperOrientationi "landscape" PaperPositionMode "auto" PaperType "usletter" PaperUnits "inches" ZoomFactor "100" AutoZoom on ReportName "simulink-default.rpt" Block { BlockType DiscretePulseGenerator Name "Discrete Pulse\nGenerator" Position [45, 25, 75, 55] Amplitude "1" Period "2" PulseWidth "1" PhaseDelay "0" SampleTime "1" } Block { BlockType Product Name "Product" Ports [2, 1, 0, 0, 0] Position [145, 67, 175, 98] Inputs "2" SaturateOnIntegerOverflow on } ... } ... [Dean, Cordy, Malton, Schneider JASE 2003]

SLIDE 31

filtering

removes more than 300 kinds

f irrelevant

elements and blocks

increases signal-to-noise

ratio in text

... System { Name "onoff” Block { BlockType DiscretePulseGenerator Name "Discrete Pulse\nGenerator” Amplitude "1" Period "2" PulseWidth "1" PhaseDelay "0" SampleTime "1" } Block { BlockType Product Name "Product" Ports [2, 1, 0, 0, 0] Inputs "2” } ... } ...

SLIDE 32

filtering significantly improved performance

precision - 10x fewer false positives hand validation of results

recall - many fewer false negatives

fewer missed clones much larger clones

but:

some clones we could clearly see by hand still not detected - why?

SLIDE 33

Challenge #4

no linear order of model elements

SLIDE 34

Challenge #4

solution: topological sort by block, line, port, branch

SLIDE 35

sorting

increases recall, to find many more clones

neutral_up_down 1 mutually_exclusive neutral up down validated_neutral validated_up validated_down check_up action reset checked_action check_down action reset checked_action Goto1 [reset] From2 [reset] From1 [reset] reset 4 down 3 up 2 neutral 1

neutral_up_down