1. AGENT SIMULATIONS ON GRAPHICS HARDWARE Timothy Johnson - twjohnson@students.latrobe.edu.au Supervisor: Dr. John Rankin 1

2. OUTLINE  Agent Simulations  Graphics Hardware  System  Model  Planning  Code Generation  Simulation  Analysis  Examples  Performance 2

3. AGENT SIMULATIONS  Simulations with hundreds or thousands of software agents  Agent-based simulations are becoming increasingly more popular [1]  Each entity simulated at the individual microscopic level  Emergence of interesting complex behaviour at the macroscopic level  Fields: air traffic control, disaster management, biology, crime analysis, economics, pedestrian dynamics, military simulations, etc  Problems  Expensive and time consuming to develop  Limited in scope by available processing power 3

4. GRAPHICS HARDWARE  Graphical Processing Units (GPUs) are outstripping Central Processing Units (CPUs) in terms of raw performance [2]  Performance gap continuing to widen  Example  Intel Core i7-4930K [$679][187.2 GFLOPS]  NVIDIA GTX 780 [$639][3977 GFLOPS]  Programming Difficulties  Dealing with hundreds or thousands of cores and threads  Data synchronisation  Memory architecture  Data transfer between CPU and GPU  Shortage of high quality learning resources Prices retrieved from pccasegear.com on 04/11/2013 4

5. IDEA  Increase processing power, allowing more agents to be simulated than previously possible  Make a system that does not require the user to perform GPU programming 5

6. SYSTEM OVERVIEW Model Code Generation Planning Simulation Analysis 6

7. SYSTEM OVERVIEW Model Code Generation Planning Simulation Analysis 7

8. MODEL 8

9.  Resources – Items agents interact with  Apple, Bread, Car, Gold, Water  Attributes – Represent aspects of agents  Age, Height, Health, Hunger, Energy  Rules – Specify how an agent can interact with the world  Move, Use, Harvest, Trade, Combine  Goals – Define the agent’s behaviour during simulation  Maximize Attributes, Maximize Resources  Agent Types – Built up with Attributes, Rules and Goals  Farmer, Duck, Bee, Policeman  Communities – Consists of one or more Agent Types  Town, Beehive, Sports team 9

10. MODEL: DUCK SIMULATION  Resources  Bugs  Bread  Attributes  Hunger  Rules  Use (Bread, Hunger, -5)  Use (Bugs, Hunger, -3)  Harvest (Bread)  Harvest (Bugs)  Goals  Minimize Attribute (Hunger) 10

11. SYSTEM OVERVIEW Model Code Generation Planning Simulation Analysis 11

12. PLANNING  User enters resources, attributes, rules, goals, agents, communities into our tool  System then automatically links goals and rules to create plans Goal One Rule One Rule Two Rule Three 12

13. PLANNING: DUCK SIMULATION 13

14. PLANNING: DUCK SIMULATION Goal Minimize Attribute (Hunger) Goal Minimize Attribute (Hunger) Rule Harvest (Bread) Rule Harvest (Bread) Rule Harvest (Bugs) Rule Harvest (Bugs) Rule Use (Bread, Hunger, -5) Rule Use (Bread, Hunger, -5) Rule Use (Bugs, Hunger, -3) Rule Use (Bugs, Hunger, -3) 14

15. SYSTEM OVERVIEW Model Code Generation Planning Simulation Analysis 15

16. CODE GENERATION  OpenCL code kernels are created for each agent type defined  OpenCL is programmed through a subset of C with its own extensions  Can run on both NVIDIA and AMD GPUs  Algorithm  For each agent type  Create a new kernel  Prepare some initial definitions and kernel arguments  Iterate through each plan created in the previous stage  Generate segments of code for each rule in each plan  By automatically generated optimised code, we manage to hide the difficulties of GPGPU from the user 16

17. CODE GENERATION: DUCK SIMULATION Goal Minimize Attribute (Hunger) Goal Minimize Attribute (Hunger) Rule Harvest (Bread) Rule Harvest (Bread) Rule Harvest (Bugs) Rule Harvest (Bugs) Rule Use (Bread, Hunger, -5) Rule Use (Bread, Hunger, -5) Rule Use (Bugs, Hunger, -3) Rule Use (Bugs, Hunger, -3) 17

18. SYSTEM OVERVIEW Model Code Generation Planning Simulation Analysis 18

19. SIMULATION  Code kernels are compiled  Data is generated on the CPU  Data is transferred across to GPU memory  Each agent to be simulated is assigned its own thread  Simulation begins to run 19

20. SIMULATION: DUCK SIMULATION 20

21. SIMULATION: DUCK SIMULATION Yellow = Ducks, Green = Bugs, Blue = Bread 21

22. SYSTEM OVERVIEW Model Code Generation Planning Simulation Analysis 22

23. ANALYSIS  All data from the simulation, including agent positions, attribute values, resource values, plans chosen, etc can be output to file during runtime  After the simulation run has completed, this data can then be analysed externally by other tools 23

24. ANALYSIS: DUCK SIMULATION 24

25. SCENARIOS Farming Simulation [3] Bee Simulation [4] Factory Simulation 25

26. PERFORMANCE TESTING  Two versions of the “Farming Simulation” created  GPU version using our system  Custom CPU-optimised version  Tests carried out on four GPUs and four CPUs  Two laptop machines and two desktop machines  Millisecond per update and frame rates were calculated  Testing found modern GPUs are faster than CPUs in most cases 26

27. SYSTEM PERFORMANCE 27

28. BEYOND SIMULATIONS  Movies  Games  Advertisements 28

29. QUESTIONS? Timothy Johnson - twjohnson@students.latrobe.edu.au Supervisor: Dr. John Rankin References: [1] Chan, W. K. V., Son Y. J., Macal, C. M., Agent-based simulation tutorial - simulation of emergent behavior and differences between agent-based simulation and discrete-event simulation, In Proceedings of the Winter Simulation Conference, WSC '10, 135-150, 2010. [2] Luebke, D., Humphreys, G., How GPUs Work, Computer, 40 (2), 96-100, 2007. [3] Johnson, T. W. C., Rankin, J. R., Parallel Agent Systems on a GPU for use with Simulations and Games, In Proceedings of the 1st International Conference on Computing, Information Systems and Communications, CISCO '01, 229-235, 2012. [4] Johnson, T. W. C., Rankin, J. R., User Friendly Agent-Oriented Simulation Builder, In Proceedings of the 8th International Conference on Information Technology and Applications, ICITA '08, 2013. 29