Affinity, locality, and load balancing in scheduling user program-level threads for execution by a computer system

• US 6,289,369 B1
• Filed: 08/25/1998
• Issued: 09/11/2001
• Est. Priority Date: 08/25/1998
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Sundaresan, Neelakantan
• Primary Examiner(s): BANANKHAH, MAJID A
• Application Number:
  US09/140,129
• Time in Patent Office: 1,113 Days
• Field of Search: 709/100, 709/101, 709/102, 709/103, 709/104, 709/105, 709/106, 709/107, 709/108, 703/27, 710/23, 711/6, 711/133, 711/148, 711/153, 712/28, 712/1
• US Class Current: 718/103
• CPC Class Codes: G06F 9/5033 considering data affinity
• CPC Class Codes: G06F 9/505 considering the load

Abstract

A technique for dynamically exploiting affinity, locality, and load balancing in scheduling the execution of multi-threaded user programs in a multi-processor computer system. Affinity, locality, and load balancing characteristics are specified at thread creation time and can be dynamically changed during thread execution, either by the user program itself or by any other process or entity with sufficient privileges and information. A central schedule queue and one or more per-processor local schedule queues are used to schedule the threads based on the dynamically changing affinity, locality, and load balancing characteristics.

Claims

1. A method of scheduling thread execution in a computer, comprising the steps of:

(a) creating a thread in a memory of a computer,

(b) specifying one or more scheduling characteristics of the thread;

(c) scheduling execution of the thread in the computer in accordance with the specified scheduling characteristics; and

(d) modifying one or more of the scheduling characteristics of the thread during the execution of the thread.

2. The method of claim 1, wherein the specified scheduling characteristics are selected from a group comprising a locality characteristic, an affinity characteristic, and a load balancing characteristic.

3. The method of claim 2, wherein the locality characteristic indicates that the thread is executed by a specified processor.

4. The method of claim 3, further comprising the step of scheduling the thread for execution using a local schedule queue of the specified processor.

5. The method of claim 2, wherein the locality characteristic indicates that the thread is local to a subset of available processors.

6. The method of claim 2, wherein the locality characteristic indicates that the thread can execute on any processor.

7. The method of claim 2, wherein the affinity characteristic indicates that the thread can reuse data from a prior thread.

8. The method of claim 2, wherein the affinity characteristic indicates that the thread is affine to another thread.

9. The method of claim 8, further comprising the step of scheduling the affine threads for execution back-to-back in a processor when the affine threads have a same locality characteristic.

10. The method of claim 8, wherein the affine threads operate on shared data.

11. Method of claim 8, further comprising the step of scheduling a second thread for execution in a processor after a first thread has completed its execution, when the first and second threads are affine threads, so that the second thread can reuse cached data in the processor.

12. The method of claim 1, wherein the specifying step comprises the step of specifying one or more of the specified scheduling characteristics of the thread during execution of the thread.

13. The method of claim 12, wherein the scheduling characteristics are specified b the thread itself.

14. The method of claim 1, wherein the specifying step comprises the step of specifying one or more of the scheduling characteristics of the thread when the thread is created.

15. The method of claim 14, wherein the scheduling characteristics are specified by the thread itself.

16. The method of claim 1, wherein the scheduling characteristics are modified by the thread itself.

17. The method of claim 1, wherein the scheduling characteristics are modified b an entity other than the thread.

18. The method of claim 1, wherein the modifying step comprises the step of dynamically modifying one or more of the scheduling characteristics of the thread during execution of the thread based on an operation of the thread.

19. The method of claim 1, wherein the modifying step comprises the step of dynamically modifying one or more of the scheduling characteristics of the thread during execution of the thread based on a data access pattern of the thread.

20. The method of claim 1, wherein the specifying step comprises the step of specifying one or more scheduling characteristics of the thread via a user-level thread library.

21. The method of claim 20, wherein the user-level thread library provides facilities to specify the affinity, locality, and load balancing characteristics used in scheduling the thread for execution.

22. The method of claim 1, wherein the scheduling step comprises the step of scheduling execution of the thread in the computer in accordance with the specified scheduling characteristics using a central schedule queue and a per-processor local schedule queue.

23. The method of claim 1, wherein the thread is non-sticky and thus is schedulable on any available processor.

24. The method of claim 1, wherein the thread is part-sticky and thus is schedulable on a set or subset of available processors.

25. The method of claim 1, wherein the thread is sticky and thus is executed only on a specified processor.

26. A multi-threaded computer system, comprising:

(a) one or more processors;

(b) means, performed by one of the processors, for creating a thread in a memory of the computer system;

(c) means, performed by one of the processors, for specifying one or more scheduling characteristics of the thread;

(d) means, performed by one of the processors, for scheduling execution of the thread in the computer in accordance with the specified scheduling characteristics; and

(e) means, performed by one of the processors, for modifying one or more of the scheduling characteristics of the thread during the execution of the thread.

27. The system of claim 26, wherein the specified scheduling characteristics are selected from a group comprising a locality characteristic, an affinity characteristic, and a load balancing characteristic.

28. The system of claim 27, wherein the locality characteristic indicates that the thread is executed by a specified processor.

29. The system of claim 28, further comprising means for scheduling the thread for execution using a local schedule queue of the specified processor.

30. The system of claim 27, wherein the locality characteristic indicates that the thread is local to a subset of available processors.

31. The system of claim 27, wherein the locality characteristic indicates that the thread is can execute on any processor.

32. The system of claim 27, wherein the affinity characteristic indicates that the thread can reuse data from a prior thread.

33. The system of claim 27, wherein the affinity characteristic indicates that the thread is affine to another thread.

34. The system of claim 33, further comprising means for scheduling the affine threads for execution back-to-back in a processor when the affine threads have a same locality characteristic.

35. The system of claim 33, wherein the affine threads operate on shared data.

36. The system of claim 33, further comprising means for scheduling a second thread for execution in a processor after a first thread has completed its execution, when the first and second threads are affine threads, so that the second thread can reuse cached data in the processor.

37. The system of claim 26, wherein the means for specifying comprises means for specifying one or more of the specified scheduling characteristics of the thread during execution of the thread.

38. The system of claim 37, wherein the scheduling characteristics are specified by the thread itself.

39. The system of claim 26, wherein the means for specifying comprises means for specifying one or more of the scheduling characteristics of the thread when the thread is created.

40. The system of claim 39, wherein the scheduling characteristics are specified by the thread itself.

41. The system of claim 26, wherein the scheduling characteristics are modified by the thread itself.

42. The system of claim 26, wherein the scheduling characteristics are modified by an entity other than the thread.

43. The system of claim 26, wherein the means for modifying comprises means for dynamically modifying one or more of the scheduling characteristics of the thread during execution of the thread based on an operation of the thread.

44. The system of claim 26, wherein the means for modifying comprises means for dynamically modifying one or more of the scheduling characteristics of the thread during execution of the thread based on a data access pattern of the thread.

45. The system of claim 26, wherein the means for specifying comprises means for specifying one or more scheduling characteristics of the thread via a user-level thread library.

46. The system of claim 45, wherein the user-level thread library provides facilities to specify the affinity, locality, and load balancing characteristics used in scheduling the thread for execution.

47. The system of claim 26, wherein the means for scheduling comprises means for scheduling execution of the thread in the computer in accordance with the specified scheduling characteristics using a central schedule queue and a per-processor local schedule queue.

48. The system of claim 26, wherein the thread is non-sticky and thus is schedulable on any available processor.

49. The system of claim 26, wherein the thread is part-sticky and thus is schedulable on a set or subset of available processors.

50. The system of claim 26, wherein the thread is sticky and thus is executed only on a specified processor.

51. A carrier embodying logic for scheduling thread execution in one or more processors, the logic comprising the steps of:

(a) creating a thread in a memory of the processor;

(b) specifying one or more scheduling characteristics of the thread;

(c) scheduling execution of the thread in one or more of the processors in accordance with the specified scheduling characteristics; and

(d) modifying one or more of the scheduling characteristics of the thread during the execution of the thread.

52. The method of claim 51, wherein the specified scheduling characteristics are selected from a group comprising a locality characteristic, an affinity characteristic, and a load balancing characteristic.

53. The method of claim 52, wherein the locality characteristic indicates that the thread is executed by a specified processor.

54. The method of claim 53, further comprising the step of scheduling the thread for execution using a local schedule queue of the specified processor.

55. The method of claim 52, wherein the locality characteristic indicates that the thread is local to a subset of available processors.

56. The method of claim 52, wherein the locality characteristic indicates that the thread is can execute on any processor.

57. The method of claim 52, wherein the affinity characteristic indicates that the thread can reuse data from a prior thread.

58. The method of claim 52, wherein the affinity characteristic indicates that the thread is affine to another thread.

59. The method of claim 58, further comprising the step of scheduling the affine threads for execution back-to-back in a processor when the affine threads have a same locality characteristic.

60. The method of claim 58, wherein the affine threads operate on shared data.

61. The method of claim 58, further comprising the step of scheduling a second thread for execution in a processor after a first thread has completed its execution, when the first and second threads are affine threads, so that the second thread can reuse cached data in the processor.

62. The method of claim 51, wherein the specifying step comprises the step of specifying one or more of the specified scheduling characteristics of the thread during execution of the thread.

63. The method of claim 62, wherein the scheduling characteristics are specified by the thread itself.

64. The method of claim 51, wherein the specifying step comprises the step of specifying one or more of the scheduling characteristics of the thread when the thread is created.

65. The method of claim 64, wherein the scheduling characteristics are specified by the thread itself.

66. The method of claim 51, wherein the scheduling characteristics are modified by the thread itself.

67. The method of claim 51, wherein the scheduling characteristics are modified by an entity other than the thread.

68. The method of claim 51, wherein the modifying step comprises the step of dynamically modifying one or more of the scheduling characteristics of the thread during execution of the thread based on an operation of the thread.

69. The method of claim 51, wherein the modifying step comprises the step of dynamically modifying one or more of the scheduling characteristics of the thread during execution of the thread based on a data access pattern of the thread.

70. The method of claim 51, wherein the specifying step comprises the step of specifying one or more scheduling characteristics of the thread via a user-level thread library.

71. The method of claim 70, wherein the user-level thread library provides facilities to specify the affinity, locality, and load balancing characteristics used in scheduling the thread for execution.

72. The method of claim 51, wherein the scheduling step comprises the step of scheduling execution of the thread in the computer in accordance with the specified scheduling characteristics using a central schedule queue and a per-processor local schedule queue.

73. The method of claim 51, wherein the thread is non-sticky and thus is schedulable on any available processor.

74. The method of claim 51, wherein the thread is part-sticky and thus is schedulable on a set or subset of available processors.

75. The method of claim 51, wherein the thread is sticky and thus is executed only on a specified processor.