1 the effect of code coverage on fault detection capability: an experimental evaluation and possible...

1

The Effect of Code Coverage on Fault Detection Capability: An Experimental Evaluation and Possible Directions

Teresa Xia Cai

Group MeetingFeb. 21, 2006

2

Outline

Testing coverage and testing strategies Research questions Experimental setup Results and analysis Discussions and conclusions

3

Introduction

Test case selection and evaluation is a key issue in software testing

Testing strategies aim to select an effective test set to detect as more faults as possible Black-box testing (functional testing) White-box testing (structural testing)

4

White-box testing schemes:Control/data flow coverage

Code coverage - measured as the fraction of program codes that are executed at least once during the test. Block coverage - the portion of basic blocks

executed. Decision coverage - the portion of decisions

executed C-Use - computational uses of a variable. P-Use - predicate uses of a variable

5

Code coverage: an indicator for test effectiveness?

Supportive empirical studies high code coverage brings high software reliability and low

fault rate both code coverage and fault detected in programs grow

over time, as testing progresses. Weyuker et al (1985, 1988, 1990) Horgan, London & Lyu (1994) Wong, Horgan, London & Mathur (1994) Frate, Garg, Mathur & Pasquini (1995)

Oppositive empirical studies Can this be attributed to causal dependency between code

coverage and defect coverage? Briand & Pfahl (2000)

6

Black-box testing schemes:testing profiles

Functional testing – based on specified functional requirements

Random testing - the structure of input domain based on a predefined distribution function

Normal operational testing – based on normal operational system status

Exceptional testing - based on exceptional system status

{

{

7


Outline

8

Research questions

1. Is code coverage a positive indicator for testing effectiveness?

2. Does such effect vary under various testing profiles?

3. Does such effect vary with different coverage measurements?

4. Is code coverage a good filter to reduce the size of effective test set?

9


Outline

10

Experimental setup

In spring of 2002, 34 teams are formed to develop a critical industry application for a 12-week long project in a software engineering course

Each team composed of 4 senior-level undergraduate students with computer science major from the Chinese University of Hong Kong

11

Experimental project description

Geometry Data flow diagram

Redundant Strapped-Down Inertial Measurement Unit (RSDIMU)

12

Software development procedure

1. Initial design document ( 3 weeks)

2. Final design document (3 weeks)

3. Initial code (1.5 weeks)

4. Code passing unit test (2 weeks)

5. Code passing integration test (1 weeks)

6. Code passing acceptance test (1.5 weeks)

13

Mutant creation

Revision control was applied in the project and code changes were analyzed

Faults found during each stage were also identified and injected into the final program of each version to create mutants

Each mutant contains one design or programming fault

426 mutants were created for 21 program versions

14

Program metricsId Lines Modules Functions Blocks Decisions C-Use P-Use Mutants

01 1628 9 70 1327 606 1012 1384 25

02 2361 11 37 1592 809 2022 1714 21

03 2331 8 51 1081 548 899 1070 17

04 1749 7 39 1183 647 646 1339 24

05 2623 7 40 2460 960 2434 1853 26

07 2918 11 35 2686 917 2815 1792 19

08 2154 9 57 1429 585 1470 1293 17

09 2161 9 56 1663 666 2022 1979 20

12 2559 8 46 1308 551 1204 1201 31

15 1849 8 47 1736 732 1645 1448 29

17 1768 9 58 1310 655 1014 1328 17

18 2177 6 69 1635 686 1138 1251 10

20 1807 9 60 1531 782 1512 1735 18

22 3253 7 68 2403 1076 2907 2335 23

24 2131 8 90 1890 706 1586 1805 9

26 4512 20 45 2144 1238 2404 4461 22

27 1455 9 21 1327 622 1114 1364 15

29 1627 8 43 1710 506 1539 833 24

31 1914 12 24 1601 827 1075 1617 23

32 1919 8 41 1807 974 1649 2132 20

33 2022 7 27 1880 1009 2574 2887 16

Average 2234.2 9.0 48.8 1700.1 766.8 1651.5 1753.4 Total: 426

15

Fault effect code lines

Lines Number Percentage

1 line: 116 27.23%

2-5 lines: 130 30.52%

6-10 lines: 61 14.32%

11-20 lines: 43 10.09%

21-50 lines: 53 12.44%

>51 lines: 23 5.40%

Average 11.39

16

Setup of evaluation test A test coverage tool was employed to analyze the com

pare testing coverage 1200 test cases were exercised on 426 mutants All the resulting failures from each mutant were analyz

ed, their coverage measured, and cross-mutant failure results compared

60 Sun machines running Solaris were involved in the test, where one cycle took 30 hours and a total of 1.6 million files around 20GB were generated

17

Test case descriptionCase ID Description of the test cases.

1 A fundamental test case to test basic functions.

2-7 Test cases checking vote control in different order.

8 General test case based on test case 1 with different display mode.

9-19 Test varying valid and boundary display mode.

20-27 Test cases for lower order bits.

28-52 Test cases for display and sensor failure.

53-85 Test random display mode and noise in calibration.

87-110 Test correct use of variable and sensitivity of the calibration procedure.

86, 111-149 Test on input, noise and edge vector failures.

150-151 Test various and large angle value.

152-392 Test cases checking for the minimal sensor noise levels for failure declaration.

393-800 Test cases with various combinations of sensors failed on input and up to one additional sensor failed in the edge vector test.

801-1000 Random test cases. Initial random seed for 1st 100 cases is: 777, for 2nd 100 cases is: 1234567890

1001-1200 Random test cases. Initial random seed is: 987654321 for 200 cases.

18

Testing coverage and testing strategies Research questions Experimental setup Results and analysis

Effective of code coverage Under various testing profiles With different coverage measurements Effective test set

Discussions and conclusions

Outline

19

Fault detection related to changes of test coverageVersion ID Blocks Decisions C-Use P-Use Any

1 6/11 6/11 6/11 7/11 7/11(63.6%)

2 9/14 9/14 9/14 10/14 10/14(71.4%)

3 4/8 4/8 3/8 4/8 4/8(50.0%)

4 7/13 8/13 8/13 8/13 8/13(61.5%)

5 7/12 7/12 5/12 7/12 7/12(58.3%)

7 5/11 5/11 5/11 5/11 5/11(45.5%)

8 1/9 2/9 2/9 2/9 2/9(22.2%)

9 7/12 7/12 7/12 7/12 7/12(58.3%)

12 10/19 17/19 11/19 17/19 18/19(94.7%)

15 6/18 6/18 6/18 6/18 6/18(33.3%)

17 5/11 5/11 5/11 5/11 5/11(45.5%)

18 5/6 5/6 5/6 5/6 5/6(83.3%)

20 9/11 10/11 8/11 10/11 10/11(90.9%)

22 12/14 12/14 12/14 12/14 12/14(85.7%)

24 5/6 5/6 5/6 5/6 5/6(83.3%)

26 2/11 4/11 4/11 4/11 4/11(36.4%)

27 4/9 5/9 4/9 5/9 5/9(55.6%)

29 10/15 10/15 11/15 10/15 12/15(80.0%)

31 7/15 7/15 7/15 7/15 8/15(53.3%)

32 3/16 4/16 5/16 5/16 5/16(31.3%)

33 7/11 7/11 9/11 10/11 10/11(90.9%)

Overall 131/252 (60.0%) 145/252 (57.5%) 137/252 (53.4%) 152/252 (60.3%) 155/252 (61.5%)

20

Relations between numbers of mutants against effective percentage of coverage

21

Cumulated defect coverage on sequence

22

Cumulated block coverage on sequence

23

Cumulated defect coverage versus block coverage

R2=0.945

24

Percentage of test case coverage

Percentage of Coverage

Blocks Decision C-Use P-Use

Average 45.86% 29.63% 35.86% 25.61%

Maximum 52.25% 35.15% 41.65% 30.45%

Minimum 32.42% 18.90% 23.43% 16.77%

25

The correlation: various test regions

Test case coverage contribution on block coverage

Test case coverage contribution on mutant coverage

I II III IV V VII II III IV V VI

26

Test cases description

I

II

IIIIV

V

VI

27




Outline

28

In various test regions

Linear modeling fitness in various test case regions

Linear regression relationship between block coverage and defect coverage in the whole test set

29

In various test regions (cont’)

Linear regression relationship between block coverage and defect coverage in region VI

Linear regression relationship between block coverage and defect coverage in region IV

30

In various test regions (cont’)

Observations: Code coverage: a moderate indicator Reasons behind the big variance between

region IV and VI

Region IV Region VI

Design principle Functional testing

Random testing

Coverage range 32% ~ 50% 48% ~ 52%

Number of exceptional test

cases

277

(Total: 373)

0

31

With functional/random testing

Code coverage:

– a moderate indicator

Random testing

– a necessary complement to functional testing Similar code coverage Both have high fault detection capability

Testing profile (size) R-square

Whole test set (1200) 0.781

Functional test cases (800)

0.837

Random test cases (400) 0.558

32

With functional/random testing (cont’)

Failure number of mutants detected only by functional testing or random testing

Test case type

Mutants detected exclusively

(total mutants killed)

Average number of test cases that

detect these mutants

Std. deviation

Functional testing

20

(382)

4.50 3.606

Random testing

9

(371)

3.67 2.236

33

Under normal operational / exceptional testing

The definition of operational status and exceptional status Defined by specification Application-dependent

For RSDIMU application Operational status: at most two sensors failed as the input

and at most one more sensor failed during the test Exceptional status: all other situations

The 1200 test cases are classified to operational and exceptional test cases according to their inputs and outputs

34

Under normal operational / exceptional testing (cont’)

Normal operational testing very weak correlation

Exceptional testing strong correlation

Testing profile (size) R-square

Whole test case (1200) 0.781

Normal testing (827) 0.045

Exceptional testing (373) 0.944

35


Normal testing: small coverage range (48%-52%) Exceptional testing: two main clusters

36


Failure number of mutants detected only by normal operational testing or exceptional testing

Test case type

Mutants detected exclusively

(total mutants detected)

Average number of test cases that detect

these mutants

Std. deviation

Normal testing

36/371 120.00 221.309

Exceptional testing

20/355 55.05 99.518

37

Under testing profile combinations

Combinations of testing profiles

Observations: Combinations containing exceptional testing indicate

strong correlations Combinations containing normal testing inherit weak

correlations

38




Outline

39

With different coverage measurements

Similar patterns as block coverage Insignificant difference under normal testing Decision and P-use have a bit larger correlation, as

they relate to change of control flow

40




Outline

41

The reduction of the test set size using coverage increase information

42


Outline

43

Answers to RQs

1. Is code coverage a positive indicator for testing effectiveness?

Our answer is supportive At most situations (61.5%), there is an coverage

increase when a test case detect additional faults. Under some functional and exceptional testing

region, the correlation between code coverage and fault coverage is pretty high

When more cumulated code coverage have been achieved, more faults are detected.

44

Answers to RQs (cont’)2. Does such effect vary under various testing

profiles? A significant correlation exists in exceptional

test cases, while no correlation in normal operational test cases.

Higher correlation is revealed in functional testing than in random testing, but the difference is insignificant

45

Answers to RQs (cont’)

3. Does such effect vary with different coverage measurements?

Not obvious with four coverage measurements

4. Is code coverage a good filter to reduce the size of effective test set?

Yes, 203 test cases (17% of the original test set) which achieve any coverage increase can detect 98% of the faults.

46

Conclusion Code coverage is a reasonably good indictor for fault

detection capability. The strong correlation revealed in exceptional testing

implies that coverage works predictably better in certain testing profiles than others.

Testing guidelines and strategy can be established for coverage-based testing: For normal operational testing: specification-based, regardless

of code coverage For exceptional testing: code coverage is an important metrics

for testing capability A quantifiable testing strategy may emerge by combining

black-box and white-box testing strategies appropriately.

1 the effect of code coverage on fault detection capability: an experimental evaluation and possible...

Documents