•Large-scalereal-lifeemergencyevacuationdrillsareoftenimpracticalorimpossibleduetomanyfactors.

•Weenvisionatoolthathelpsplannersandsituationcommandersinteractivelycreateandverifylarge-scaleemergencyresponseplans.

•Itmustbescalable,accurateandsimulatesfasterthanreal-timetoenableinteractivity.Itmustalsobeabletosimultaneouslytestformultiplecriteriaandplanswiththeabilitytocomparetheireffectivenessandhighlightpotentialpitfalls.

•ThetoolusesanetworkofGPU-enabledcomputersforsimulation,andacentralhostthatmanagestheinformationflowbetweentheseprocesses.Anexternalclientcanconnecttothehostinordertodispatchscenariodataandobtainsimulateddataforanalysis.

High-Performance Pedestrian Multi-Simulation Using GPU ClusterTwin Karmakharm, Paul Richmond

Department of Automatic Control and Systems Engineering, University of Sheffield

t.karmakharm@sheffield.ac.uk, p.richmond@sheffield.ac.uk

Pedestrian Simulation as aReal-Time Decision Support Tool

Built on the Agent-Based FLAME GPU Framework

The Model

Faster Than Real-Time

Simulation Batching

Distributed GPU Processing of Multiple Simulations

Results & Conclusion

•AgentBasedModelling(ABM)isapowerfulsimulationtechniquewhichisusedtoassessgroupbehaviourfromanumberofsimpleinteractingrulesbetweencommunicatingautonomousagents.

•SimulationcodeisgeneratedbasedontheFlexibleLarge-scaleAgentModellingEnvironmentontheGPU(FLAMEGPU).[5]

•XMLisusedformodelspecificationwithextendibleSchemasusedtoensurecorrectmodelsyntaxandaddGPUspecificinformation.Themodelspecification,alongwiththeC-basedfunctionscriptsaretransformedusingXSLTtemplatesintocompilablecode.[6,7]

•ComplexitiesofGPUprogrammingareentirelyabstractedfromthemodellingandsimulationprocess.

•Usesthesocialforceswithcontactforcemodelforsimulatinginteractionsbetweenpedestriansandtheirenvironment.[1,2]

•Pedestrianshaveaccesstoenvironmentcollisiondetectionandglobalnavigationthroughagridofnavigationagentsthatholdsasetofvectorforcefields.[3]

•Acollisionvectorforcefieldholdsinformationaboutstaticenvironments(e.g.walls)

•Navigationvectorforcefieldsrepresentthegoalsoftheagent.Itgivestheshortestpathtoasingleexit(usedfornormalwalkingmodels),multipledestinations(evacuations),oragatheringpoint.

•Industry-standardmetricsareusedtocollectdataforanalysissuchastheFruinLevelofService.[4]

•Thesimulationcanrunfasterthanreal-timewhenvisualisationisnotrequired.

•Largertime-stepscanbeusedinordertoincreasethesimulationrate,withsomeaccuracytrade-offs.

•Smallermodelsaregroupedtogetherintoasinglesimulation.Efficiencyisincreasedforlargersimulationsizesduetoreducedkernelcallsandensuringmaximumthreadutilization.

•ThegridofnavigationagentsarespatiallytranslatedtoitsglobalcoordinateaccordingtotheirsimulationID.

•AgentsareassignedIDsinordertopreventinterferenceinboundarycases.

•Amastercomputerhandlestheschedulinganddistributionofsimulations.

•Aninstanceofaslaveschedulerispresentoneachsimulationcomputerandisresponsibleforinitialisinginstancesofsimulatorprograms.

•AninstanceofasimulatoriscreatedforeachGPUcoreonthesystem.Thisavoidscomplicationsinprogramdesignasthereisnocross-communicationbetweenthesimulators.

•Simulatorsstoretheirsimulationresultslocallyandtransmitthemwhenrequested.

•Resultsarecollectedasperuser-specifiedsimulationtimeperiod.Performancecanbefurtherincreasedbyincreasingdatacollectionperiod.

•Simulationof160,000pedestriansat~20framespersecondwithoutvisualisationwhenrunning8simultaneoussimulationof2562gridsizeacross2NvidiaGTX590GPUs.

•Useslow-costhardwareandnetworkingcomponents.

•Futurework:Furtheroptimizationofthesimulationframework,incorporatingmorecomplexnavigationsystemandbringingthesoftwaretocommercialisation.

Simulation

Code

XSLT

Simulation

Templates

■■■

XSLT

Processor

■■■

Simulation Program

XML Schemas

XML Model File

Scripted

Behaviour

Simulator PC 1

Client PC 1

Simulator Manager

Node

ClientNode

Host PC

MasterScheduler

GPUSimulatorNode

GPUSimulatorNode

Navigation Agents

Col

lisio

n FV

FN

avig

atio

n FV

F 1

Nav

igat

ion

FVF

2

Navigation Agents Grid

x

y

Destination/ExitStatic Obstacle

Collision Force VectorNavigation Force Vector

Navigation Agents Pedestrian List

Global Navigation MapGlobal Pedestrian List

Simulation 1

Simulator

Navigation Agents Pedestrian List

Simulation 2

1 2 3

4 5

987

6

1 2 3

4 5

987

6

1 2 3

7 8

151413

9

4 5 6

10 11

181716

12

22 23 24

28 29

363534

30

19 20 21

25 26

333231

27

0.46 m2

0.93 m2

1.39 m2

2.32 m2

3.25 m2

3.72 m2

LOS F

LOS E

LOS D

LOS C

LOS B

LOS A

[1] D. Helbing. A mathematical model for the behavior of pedestrians. Behavioral Science, 36(4):298{310, Oct. 1991.[2] D. Helbing, I. Farkas, and T. Vicsek. Simulating dynamical features of escape panic. Nature, 407(6803):487{490, Sept. 2000.[3] T. Karmakharm, P. Richmond, and D. Romano. Agent-based Large Scale Simulation of Pedestrians With Adaptive Re-alistic Navigation Vector Fields. In Theory and Practice of Computer Graphics (TPCG) 2010, pages 67{74, 2010.[4] J. J. Fruin. Pedestrian planning and design. Metropolitan Association of Urban Designers and Environmental Plan-ners, 1971.

[5] FLAME GPU Website, http://www.flamegpu.com[6] P. Richmond and D. Romano. Agent Based GPU, a Real-time 3D Simulation and Interactive Visualisation Framework for Massive Agent Based Modelling on the GPU. The rst International Workshop on Super Visualization (IWSV08), 2008.1[7] P. Richmond and D. Romano. Template-Driven Agent-Based Modeling and Simulation with CUDA. In W.-m. W. Hwu, editor, GPU Computing Gems Emerald Edition, Applications of GPU Computing Series, chapter 21, pages 313-324. Mor-gan Kaufmann, 1st edition, Feb. 2011.

References

Work funded by BAE Systems

GPU Computing Gems Emeald Edition: Chapter 21 p.313-324