Characterizing the Energy Efficiency of Java’s Thread-Safe Collections in a Multi-Core Environment

Transcript

1. Characterizing the Energy Efficiency of Java’s Thread-Safe Collections in a Multi-Core Environment Gustavo Pinto ghlp@cin.ufpe.br Fernando Castor castor@cin.ufpe.br
2. Motivation (1/3) • First, energy consumption is a concern for unwired devices and also for data centers • Second, there is a large body of work in hardware/architecture, OS, runtime systems • However, little is known about the application level 2
3. Motivation (2/3) 3 • First, multicore CPUs are ubiquitous • Second, more cores used more power consumed • However, little is known about the energy-efficiency of multicore programs
4. Motivation (3/3) 4 • Data structures are the fundamentals of computer programming
5. 5 List<Object> lists = …; • ArrayList • LinkedList
6. 6 List<Object> lists = new ArrayList<>();
7. 7 List<Object> lists = new ArrayList<>(); Thread
8. 8 • ArrayList • LinkedList • Vector • Collections.synchronizedList() • CopyOnWriteArrayList List<Object> lists = …;
9. List<Object> lists = new Vector<>(); 9
10. 10 List<Object> lists = new Vector<>(); Thread
11. 11 List<Object> lists = new Vector<>(); Thread Thread-safe!
12. 12 List<Object> lists = new Vector<>(); Thread { } … Thread-safe!
13. 13 List<Object> lists = new Vector<>(); Thread { } … Thread-safe!
14. List<Object> lists = new CopyOnWriteArrayList<>(); 14
15. List<Object> lists = new CopyOnWriteArrayList<>(); 15 Thread
16. List<Object> lists = new CopyOnWriteArrayList<>(); 16 Thread Thread-safe!
17. List<Object> lists = new CopyOnWriteArrayList<>(); 17 { } Thread Thread-safe!
18. List<Object> lists = new CopyOnWriteArrayList<>(); 18 { } Thread Thread-safe!
19. List<Object> lists = new CopyOnWriteArrayList<>(); 19 { } Thread Thread-safe!
20. 20 • ArrayList • LinkedList • …. • Vector • Collections.synchronizedList() • CopyOnWriteArrayList • …. List<Object> lists = …;
21. 21 List<Object> lists = …; Set<Objects> sets = …; Map<Object, Object> maps = …;
22. 22 List<Object> lists = …; Set<Objects> sets = …; Map<Object, Object> maps = …;
23. Research Questions 23 • RQ1: Do different implementations of the same collection have different impacts on energy consumption? • RQ2: Do different operations in the same implementation of a collection consume energy differently?
24. 24 List Set Map ArrayList LinkedHashSet LinkedHashMap Vector ——— Hashtable Collections.syncList() Collections.syncSet() Collections.syncMap() CopyOnWriteArrayList CopyOnWriteArraySet ——— ——— ConcurrentSkipListSet ConcurrentSkipListMap ——— ConcurrentHashSet ConcurrentHashMap ——— ConcurrentHashSetV8 ConcurrentHashMapV8 16 Benchmarks
25. 25 List Set Map ArrayList LinkedHashSet LinkedHashMap Vector ——— Hashtable Collections.syncList() Collections.syncSet() Collections.syncMap() CopyOnWriteArrayList CopyOnWriteArraySet ——— ——— ConcurrentSkipListSet ConcurrentSkipListMap ——— ConcurrentHashSet ConcurrentHashMap ——— ConcurrentHashSetV8 ConcurrentHashMapV8 16 Benchmarks Not-thread safe Thread safe
26. 26 List Set Map ArrayList LinkedHashSet LinkedHashMap Vector ——— Hashtable Collections.syncList() Collections.syncSet() Collections.syncMap() CopyOnWriteArrayList CopyOnWriteArraySet ——— ——— ConcurrentSkipListSet ConcurrentSkipListMap ——— ConcurrentHashSet ConcurrentHashMap ——— ConcurrentHashSetV8 ConcurrentHashMapV8 16 Benchmarks x 3 Operations Traversal Insertion Removal
27. 27 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of memory, JDK version 1.7.0 11, build 21.
28. 28 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of memory, JDK version 1.7.0 11, build 21.
29. 29 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of memory, JDK version 1.7.0 11, build 21.
30. 30 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of DDR3 1600 memory, and JDK version 1.7.0 11, build 21.
31. List 31 Time
32. List 32 Time
33. List 33 Time 1.98x more energy!
34. List 34 Time
35. List 35 Time No sync!
36. List 36 Time CopyOnWriteArrayList: +152x
37. Set 37 Time
38. Set 38 Time Time is proportional to energy Time is not proportional to energy
39. Map 39 Time
40. Map 40 Time Time is proportional to energy Time is not proportional to energy
41. Map 41 Time
42. Map 42 Time
43. Map “Tuning Knobs” 43 Initial capacity
44. 44 Load factor Map “Tuning Knobs”
45. 45 Collision 100% Map “Tuning Knobs”
47. Future Work 47 • Perform the “removal” operations on Lists • Vary the number of threads accessing the data structure • Perform the experiments in another machine
48. Characterizing the Energy Efficiency of Java’s Thread-Safe Collections in a Multi-Core Environment Gustavo Pinto ghlp@cin.ufpe.br Fernando Castor castor@cin.ufpe.br