Slides from my presentation of a paper with the same title at ParCo 2007 - http://www.fz-juelich.de/nic-series/volume38/knafla.pdf . Future computer games need parallel programming to meet their ever growing hunger for per- formance. We report on our experiences in parallelizing the game-like C++ application Open- SteerDemo with OpenMP. To enable deterministic data-parallel processing of real-time agent steering behaviour, we had to change the high-level design, and refactor interfaces for explicit shared resource access. Our experience is summarized in a set of guidelines to help parallelizing legacy game code.

1. Parallelizing
a Real-Time
Steering Simulation
for Computer Games
with
OpenMP
Bjoern Knafla and Claudia Leopold
University of Kassel (Germany)
Dienstag, 11. September 2007 1

2. Dienstag, 11. September 2007 2

3. Steering a flock of birds
Dienstag, 11. September 2007 2

4. Dienstag, 11. September 2007 3

5. Dienstag, 11. September 2007 3

6. Dienstag, 11. September 2007 3

7. Bird
Dienstag, 11. September 2007 3

8. Bird
Boid
Dienstag, 11. September 2007 3

9. Bird
Boid
Agent
Dienstag, 11. September 2007 3

10. Dienstag, 11. September 2007 4

11. CPU
core
Dienstag, 11. September 2007 4

12. Dienstag, 11. September 2007 4

13. CPU CPU CPU CPU
core core core core
CPU CPU CPU CPU
core core core core
CPU CPU CPU CPU
core core core core
CPU CPU CPU CPU
core core core core
Dienstag, 11. September 2007 4

14. Dienstag, 11. September 2007 4

15. Outline
Dienstag, 11. September 2007 5

16. Outline
1. Steering behaviors
Dienstag, 11. September 2007 5

17. Outline
1. Steering behaviors
2. OpenSteerDemo
Dienstag, 11. September 2007 5

18. Outline
1. Steering behaviors
2. OpenSteerDemo
3. Bad parallelization
Dienstag, 11. September 2007 5

19. Outline
1. Steering behaviors
2. OpenSteerDemo
3. Bad parallelization
4. Parallelization that works
Dienstag, 11. September 2007 5

20. Outline
1. Steering behaviors
2. OpenSteerDemo
3. Bad parallelization
4. Parallelization that works
5. Performance
Dienstag, 11. September 2007 5

21. Outline
1. Steering behaviors
2. OpenSteerDemo
3. Bad parallelization
4. Parallelization that works
5. Performance
6. Summary
Dienstag, 11. September 2007 5

22. Steering behaviors
Dienstag, 11. September 2007 6

23. Dienstag, 11. September 2007 7

24. Dienstag, 11. September 2007 7

25. Dienstag, 11. September 2007 7

26. Dienstag, 11. September 2007 7

27. Alignment
Dienstag, 11. September 2007 8

28. OpenSteerDemo
Dienstag, 11. September 2007 9

29. OpenSteerDemo
Testbed for the open source library
OpenSteer from Craig W. Reynods
Dienstag, 11. September 2007 9

30. OpenSteerDemo
Testbed for the open source library
OpenSteer from Craig W. Reynods
Simulates steering behavior of agents
Dienstag, 11. September 2007 9

31. OpenSteerDemo
Testbed for the open source library
OpenSteer from Craig W. Reynods
Simulates steering behavior of agents
Game-like C++ real-time application
Dienstag, 11. September 2007 9

32. OpenSteerDemo
Testbed for the open source library
OpenSteer from Craig W. Reynods
Simulates steering behavior of agents
Game-like C++ real-time application
OpenGL
Dienstag, 11. September 2007 9

33. OpenSteerDemo
Testbed for the open source library
OpenSteer from Craig W. Reynods
Simulates steering behavior of agents
Game-like C++ real-time application
OpenGL
http://opensteer.sourceforge.net
Dienstag, 11. September 2007 9

34. Main loop
Dienstag, 11. September 2007 10

35. Main loop
Input Update Output
Dienstag, 11. September 2007 10

36. Main loop
Graph-
Input Update
ics
Dienstag, 11. September 2007 11

37. Input stage
Graph-
Input Update
ics
Time step t
Dienstag, 11. September 2007 12

38. Update stage
Graph-
Input Update
ics
Time step t
Dienstag, 11. September 2007 13

39. Graphics stage
Graph-
Input Update
ics
Time step t
Dienstag, 11. September 2007 14

40. Next main loop cycle
Graph-
Input Update
ics
Time step t+1
Dienstag, 11. September 2007 15

41. Update stage
Graph-
Input Update
ics
Dienstag, 11. September 2007 16

42. Agent update
Update
agent
read + write read + write read + write write
Agent
Neighbors Random. Graphics
state
references
Dienstag, 11. September 2007 17

43. Time step t
In state t-1
Update Update Update Update
agent agent agent ... agent
0 1 2 n
Sequential
agent
update
order
Dienstag, 11. September 2007 18

44. Time step t
In state t-1
Update Update Update Update
agent agent agent ... agent
0 1 2 n
Sequential Neighbors
agent
update
order In state t-1
Dienstag, 11. September 2007 18

45. Time step t
In state t-1
Update Update Update Update
agent agent agent ... agent
0 1 2 n
Sequential
agent
update
order
Dienstag, 11. September 2007 19

46. Time step t
In state t In state t-1
Update Update Update Update
agent agent agent ... agent
0 1 2 n
Sequential
agent
update
order
Dienstag, 11. September 2007 20

47. Time step t
In state t In state t-1
Update Update Update Update
agent agent agent ... agent
0 1 2 n
Sequential
agent
update
order
Dienstag, 11. September 2007 21

48. Time step t
In state t In state t-1
Update Update Update Update
agent agent agent ... agent
0 1 2 n
Sequential Neighbors
agent
update
order In state ?
Dienstag, 11. September 2007 21

49. Bad parallelization
Dienstag, 11. September 2007 22

50. Dienstag, 11. September 2007 23

51. Questionable reliability
and correctness
Dienstag, 11. September 2007 23

52. No speedup
Dienstag, 11. September 2007 24

53. No speedup
Dienstag, 11. September 2007 24

54. Main
parallelization problems
Dienstag, 11. September 2007 25

55. Dienstag, 11. September 2007 26

56. Race conditions
Dienstag, 11. September 2007 26

57. Race conditions
Global variables, deep inheritance
hierarchies, strongly interdependent
classes
Dienstag, 11. September 2007 26

58. Race conditions
Global variables, deep inheritance
hierarchies, strongly interdependent
classes
Non-determinism
Dienstag, 11. September 2007 26

59. Race conditions
Global variables, deep inheritance
hierarchies, strongly interdependent
classes
Non-determinism
Order of agent updates and random
numbers
Dienstag, 11. September 2007 26

60. Race conditions
Global variables, deep inheritance
hierarchies, strongly interdependent
classes
Non-determinism
Order of agent updates and random
numbers
Non-thread-safe functions
Dienstag, 11. September 2007 26

61. Parallelization
that works
Dienstag, 11. September 2007 27

62. Guiding ideas
Dienstag, 11. September 2007 28

63. Guiding ideas
Update order independent simulation
Dienstag, 11. September 2007 28

64. Guiding ideas
Update order independent simulation
Agent modification only dependent on its
steering vector
Dienstag, 11. September 2007 28

65. Guiding ideas
Update order independent simulation
Agent modification only dependent on its
steering vector
Minimize synchronization
Dienstag, 11. September 2007 28

66. Guiding ideas
Update order independent simulation
Agent modification only dependent on its
steering vector
Minimize synchronization
Finish parallel processing before graphics
stage
Dienstag, 11. September 2007 28

67. Guiding ideas
Update order independent simulation
Agent modification only dependent on its
steering vector
Minimize synchronization
Finish parallel processing before graphics
stage
Interfaces for explicit context
Dienstag, 11. September 2007 28

68. Refactorization
Dienstag, 11. September 2007 29

69. Split agent update
Agent update
Sim. Modify
agent agent
Agent
update
order
Dienstag, 11. September 2007 30

70. Simulate agent stage
read + write Sim.
agent
Steering
Vector read read write
Agent Render-
Random. Neighbors
state Feeder
Agent Agent public
state references
Agent private
Dienstag, 11. September 2007 31

71. Simulate agent stage
read + write Sim.
agent
Steering
Vector read
on st read
on st write
c c
Agent Render-
Random. Neighbors
state Feeder
Agent Agent public
state references
Agent private
Dienstag, 11. September 2007 31

72. Modify agent stage
read Modify
agent
Steering
Vector read + write
Agent Render-
Random. Neighbors
state Feeder
Agent Agent public
state references
Agent private
Dienstag, 11. September 2007 32

73. Modify neighbor data
structure
Modify
neigh-
bors
Steering
Vector read write
Agent Render-
Random. Neighbors
state Feeder
Agent Agent public
state references
Agent private
Dienstag, 11. September 2007 33

74. Refactor update stage
Input Update Output
Dienstag, 11. September 2007 34

75. Split update stage
Update stage
Sim. Modify
Sub-stages: simulation and modification
Dienstag, 11. September 2007 35

76. Old update stage
Update agent Update agent Update agent
Sim. Modify Sim. Modify Sim. Modify
agent agent agent agent ... agent agent
0 0 1 1 n n
Sequential
processing
order
Dienstag, 11. September 2007 36

77. Time step t
Agents in state t-1
Sim. Modify Sim. Modify Sim. Modify
agent agent agent agent ... agent agent
0 0 1 1 n n
Sequential
processing
order
Dienstag, 11. September 2007 37

78. Time step t
Agent in state t Agents in state t-1
Sim. Modify Sim. Modify Sim. Modify
agent agent agent agent ... agent agent
0 0 1 1 n n
Sequential
processing
order
Dienstag, 11. September 2007 38

79. New update stage
Simulation sub-stage Modification sub-stage
Sim. Sim. Sim. Modify Modify Modify
agent agent ... agent agent agent ... agent
0 1 n 0 1 n
Sequential
processing
order
Dienstag, 11. September 2007 39

80. Time step t
Agents in public state t-1
Sim. Sim. Sim. Modify Modify Modify
agent agent ... agent agent agent ... agent
0 1 n 0 1 n
Sequential
processing
order
Dienstag, 11. September 2007 40

81. Time step t
Agents in public state t-1
Sim. Sim. Sim. Modify Modify Modify
agent agent ... agent agent agent ... agent
0 1 n 0 1 n
Sequential
processing
order
Dienstag, 11. September 2007 41

82. Time step t
Agents in public state t-1
Sim. Sim. Sim. Modify Modify Modify
agent agent ... agent agent agent ... agent
0 1 n 0 1 n
Sequential
processing
order
Dienstag, 11. September 2007 42

83. Time step t
Agents in public state t-1
Sim. Sim. Sim. Modify Modify Modify
agent agent ... agent agent agent ... agent
0 1 n 0 1 n
Sequential Each agent is modified
processing based solely on its own
order state
Dienstag, 11. September 2007 42

84. Parallelization
Dienstag, 11. September 2007 43

85. Parallel update stage
Simulation Modification
S S M M M
P P
B B
a S S
a M M M
a a
r r
r r Update
a S S S
a M M M
r r neigh-
l l
i i bors
l S S
l M M M
e e
e e Possibly
r r
l S S S
l M M M parallel
Dienstag, 11. September 2007 44

86. Parallel simulation
sub-stage
Simulation Modification
S S
P P
B B
a S S
a
a a
r r
r r Update
a S S S
a
r r neigh-
l l
i i bors
l S S
l
e e
e e Possibly
r r
l S S S
l parallel
Dienstag, 11. September 2007 45

87. Parallel modification
sub-stage
Simulation Modification
M M M
P P
B B
a a M M M
a a
r r
r r Update
a a M M M
r r neigh-
l l
i i bors
l l M M M
e e
e e Possibly
r r
l l M M M parallel
Dienstag, 11. September 2007 46

88. Parallel modification
sub-stage
Simulation Modification
P P
B B
a a
a a
r r
r r Update
a a
r r neigh-
l l
i i bors
l l
e e
e e Possibly
r r
l l parallel
Dienstag, 11. September 2007 47

89. Performance
Dienstag, 11. September 2007 48

90. Test-computer
Dual-processor dual-
core 2 GHz AMD
Opteron with 2 GB Ram
2x Nvidia 7800 GTX
graphics cards in SLI
mode
Linux OS
Dienstag, 11. September 2007 49

91. Max speedups
Dienstag, 11. September 2007 50

92. Max speedups
Comparison of OpenSteerDemo flock of
birds simulation with 4 threads with
OpenMP-disabled version
Dienstag, 11. September 2007 50

93. Max speedups
Comparison of OpenSteerDemo flock of
birds simulation with 4 threads with
OpenMP-disabled version
Whole application: ! 2.84
Dienstag, 11. September 2007 50

94. Max speedups
Comparison of OpenSteerDemo flock of
birds simulation with 4 threads with
OpenMP-disabled version
Whole application: ! 2.84
Update stage: !! ! 3.54
Dienstag, 11. September 2007 50

95. Summary
Dienstag, 11. September 2007 51

96. Dienstag, 11. September 2007 52

97. Refactored first
Dienstag, 11. September 2007 52

98. Refactored first
Simplicity!
Dienstag, 11. September 2007 52

99. Refactored first
Simplicity!
High-level design to min. synchronization
Dienstag, 11. September 2007 52

100. Refactored first
Simplicity!
High-level design to min. synchronization
Explicit context
Dienstag, 11. September 2007 52

101. Refactored first
Simplicity!
High-level design to min. synchronization
Explicit context
Deferred computation
Dienstag, 11. September 2007 52

102. Refactored first
Simplicity!
High-level design to min. synchronization
Explicit context
Deferred computation
Identified slices for data-parallel
computation
Dienstag, 11. September 2007 52

103. Thank you!
Dienstag, 11. September 2007 53