Recomendados
Más contenido relacionado
La actualidad más candente
La actualidad más candente (20)
Similar a ICSE2014
Similar a ICSE2014 (20)
Último
Último (20)
ICSE2014
- 1. Detec%ng Performance An%-‐pa2erns for Applica%ons Developed Using Object-‐Rela%onal Mapping 1 Mohamed Nasser, Parminder Flora Tse-‐Hsun(Peter) Chen Ahmed E. Hassan Weiyi Shang Zhen Ming Jiang
- 2. Databases are essen%al in large-‐scale soBware systems 2 Database
- 3. 3 Applica%on developers work with objects More intui*ve if we can map objects directly to DB
- 4. 4 Object-‐Rela%onal Mapping eliminates the gap between objects and SQL Database • Lots of boilerplate code • Need to manage object-‐DB transla%ons manually Object Classes Problem of using raw SQLs ORM Much less code and shorter development %me
- 5. ORM is widely used in prac%ce 5 • Java Hibernate has more than 8 million downloads • In 2013, 15% of the 17,000 Java developer jobs require ORM experience (dice.com) Different ORM technologies
- 6. An example class with ORM code 6 @En%ty @Table(name = “user”) public class User{ @Column(name=“id”) private int id; @Column(name=“name”) String userName; @OneToMany(fetch=FetchType.EAGER) List<Team> teams; public void setName(String n){ userName = n } … other ge]er and se]er methods User.java User class is mapped to “user” table in DB id is mapped to the column “id” in the user table A user can belong to mul%ple teams Eagerly retrieve associated teams when retrieving a user object
- 7. Accessing the database using ORM 7 User u = findUserByID(1); ORM database select u from user where u.id = 1; u.setName(“Peter”); update user set name=“Peter” where user.id = 1; Objects SQLs
- 8. Developers may not be aware of database access Wow! I don’t need to worry about DB code! ORM code with performance ang-‐pa]erns 8 Bad system performance The performance difference can be LARGE!
- 9. Performance an%-‐pa2ern detec%on framework Performance ang-‐pa]ern detecgon and ranking framework Ranked according to performance impact Ranked Performance ang-‐pa]erns Source Code detecgon ranking 9
- 10. Performance an%-‐pa2ern detec%on framework Performance ang-‐pa]ern detecgon and ranking framework Ranked according to performance impact Ranked Performance ang-‐pa]erns Source Code detecgon ranking 10 Discuss only one an%-‐pa2ern in the presenta%on (more pa2erns in the paper)
- 11. ORM excessive data an%-‐pa2ern Class User{ @EAGER List<Team> teams; } User u = findUserById(1); u.getName(); EOF 11 Objects SQL Eagerly retrieve teams from DB User Table Team Table join Team data is never used!
- 12. Detec%ng excessive data using sta%c analysis 12 First find all the objects that eagerly retrieve data from DB Class User{ @EAGER List<Team> teams; } Idengfy all the data usages of objects User user = findUserByID(1); Check if the retrieved data is ever used user.getName(); user team user team
- 13. Performance an%-‐pa2ern detec%on framework Performance ang-‐pa]ern detecgon and ranking framework Ranked according to performance impact Ranked Performance ang-‐pa]erns Source Code detecgon ranking 13
- 14. Performance an%-‐pa2ern detec%on framework Performance ang-‐pa]ern detecgon and ranking framework Ranked according to performance impact Ranked Performance ang-‐pa]erns Source Code detecgon ranking 14
- 15. Performance an%-‐pa2erns have different impacts 15 User user_in_1_team = findUserByID(1); Retrieving 1 user and 1 team User user_in_100_teams = findUserByID(100); Retrieving 1 user and 100 teams! One can only reveal performance impact by execu%on
- 16. Measuring the impact using repeated measurements and effect sizes 16 We use effect sizes (Cohen’s D) to measure the performance impact Effect sizes = We repeat each test 30 gmes to obtain stable measurement results Size of performance impact is not defined: Performance measurements are unstable:
- 17. Studied systems and detec%on results Large open-‐source e-‐commence system > 1,700 files > 206K LOC Enterprise system > 3,000 files > 300K LOC Spring open-‐source system Online system for a pet clinic 51 files 3.3K LOC 482 excessive data > 10 excessive data 10 excessive data 17
- 18. Performance impact Research ques%ons Ranks of the ang-‐pa]erns at different scales 18
- 19. Performance impact Research ques%ons Ranks of the ang-‐pa]erns at different scales 19
- 20. Assessing an%-‐pa2ern impact by fixing the an%-‐pa2erns Execugon Response %me user.getName() Code with ang-‐pa]erns fetchType.set(LAZY) user.getName() Code without ang-‐pa]erns 20 Execute test suite 30 gmes Response %me aOer fixing the anS-‐paTerns Avg. % improvement and effect sizes Execugon Execute test suite 30 gmes
- 21. Performance an%-‐pa2erns have medium to large effect sizes 0% 20% 40% 60% 80% 100% Excessive Data BL EA PC 21 % improvement in response gme large effect size large effect size medium effect size
- 22. Performance impact Research ques%ons Ranks of the ang-‐pa]erns at different scales 22 Removing an%-‐pa2ern improves response by ~35%
- 23. Performance impact Research ques%ons Ranks of the ang-‐pa]erns at different scales 23 Removing an%-‐pa2ern improves response by ~35%
- 24. Performance problems usually arise under large load 24
- 25. Performance problems revealed at small scales may be more serious 25 We should first fix the an%-‐pa2erns that have larger effects at smaller scales Input scales may have exponenSal effects on performance Different input scales Performance at different input scales
- 26. Comparing ranked an%-‐pa2erns at different data scales 26 Ranked Performance ang-‐pa]erns from small data detecgon ranking Ranked Performance ang-‐pa]erns from large data ? Small size input Large size input
- 27. An%-‐pa2erns have large effects on performance even at smaller data scales 27 0 10 20 30 40 50 0 1 2 3 4 5 Effect size Effect sizes and the ranks of the an%-‐pa2erns are consistent in different data scales
- 28. Performance impact Research ques%ons Ranks of the ang-‐pa]erns at different scales 28 Removing an%-‐pa2ern improves response by ~35% Ranks of the an%-‐pa2erns are consistent in different data scales
- 29. 29
- 30. 30
- 31. 31
- 32. 32