Actas JP2011 - Universidad de La Laguna

Actas XXII Jornadas de Paralelismo (JP2011) , La Laguna, Tenerife, 7-9 septiembre 2011A Scalable Visualization System for CrowdSimulationsGuillermo Vigueras, Juan M. Orduña, Miguel Lozano y Víctor Fernández-Bauset 1Abstract— The visualization system of large-scalecrowd simulations should scale up with both the numberof visuals (views of the virtual world) and thenumber of agents displayed in each visual. Otherwise,we could have large scale crowd simulationswhere only a small percentage of the population isdisplayed. Several approaches have been proposed inorder to efficiently render crowds of animated characters.However, these approaches either render crowdsanimated with simple behaviors or they can only supporta few hundreds of user-driven entities. In thispaper, we propose a distributed visualization systemfor large crowds of autonomous agents that allows thevisualization of the crowd without adding significantoverhead to the simulation servers. The proposed implementationcan be hosted on dedicated computersdifferent from the servers, and it takes advantage ofthe Graphics Processor Unit (GPU) capabilities. Asa result, the performance evaluation shows that thousandsof agents can be rendered without affecting theperformance of the simulation servers. These resultssuggest that the design of the visual client allows toadd multiple visuals for displaying large crowds.Keywords— Distributed simulation, parallel renderingI. IntroductionLA rge scale crowd simulations are becoming essentialtools for many virtual environment applicationsin education, training, and entertainment [1],[2]. In order to deal with the computational complexityof large scale simulations, different proposals havebeen made for achieving both very populated scenes[3] and scalable autonomous behaviors [4], [5]. However,the scalability of autonomous complex agents(crowd simulations) is still an open issue in spite ofthese efforts.In previous works, we proposed a distributed systemarchitecture that can simulate large crowds ofautonomous agents at interactive rates [6], [7], [8],and it can take advantage of the inherent scalabilityof manycore computer architectures [9]. However, inorder to make a truly scalable system for crowd simulations,the visualization system (the module responsibleof rendering the images of the virtual world)should also be addressed. The visualization systemshould scale up with both the number of visuals(cameras focusing on the virtual world) and the numberof agents displayed in each camera. Otherwise,we could have large scale crowd simulations whereonly a small percentage of the population could berendered (displayed).In this paper, we propose a distributed visualizationsystem that allows the visualization of the virtualworld without adding significant overhead to thesimulation servers, regardless of both the number of1 Dpto. de Informática, Univ. de Valencia e-mail:{Guillermo.Vigueras,Juan.Orduna}@uv.esvisuals and the number of agents rendered by eachvisual. In order to achieve this goal, the visualizationsystem consists of a visual client process (VCP)for each camera, and each VCP is hosted on a computerdifferent from the ones hosting the simulationservers. In this way, the connection of the visualclient does not significantly affects the performanceof the simulation system. The proposed implementationmigrates different rendering tasks of the VCPfrom the CPU to the Graphics Processor Unit (GPU)of the hosting computer, reducing the CPU workloadof the visual client and increasing throughput. Also,we use skinned instancing for reducing the renderingworkload. As a result, the performance evaluationshows that thousands of agents can be renderedwithout affecting the performance of the simulationservers. These results suggest that the design of thevisual client allows to add multiple visuals for displayinglarge crowds.The rest of the paper is organized as follows: sectionII shows some related work about visual clientsfor crowd simulations. Section III briefly describesthe distributed architecture for crowd simulationthat was previously proposed, and it shows the scalabilityproblems arising when connecting a VCP tothis kind of systems. Next, section IV describes theproposed implementation for the distributed visualclient, and section V shows the performance evaluationof the proposed implementation. Finally, sectionVI shows some conclusions.II. Related WorkFrom the graphics community, several approacheshave been proposed in order to efficiently rendercrowds of animated characters. Image-based [10] andPoint-based [11] techniques obtain interactive framerates when rendering crowded animated scenes byreducing the geometrical complexity of the 3D charactersmeshes. Other approaches use efficient parallelgraphic techniques to provide interactive graphicsperformance for crowded scenes [12]. Although thesegraphic-based approaches obtain good frame rates,they are not focused on providing scalable architectures.Other proposals [13] combine parallel architectureswith efficient graphic techniques to simulateand to display thousands of individuals. In this case,authors use the Cell processor architecture withoutconsidering scalability issues.From the distributed simulation arena, there havebeen several approaches oriented to handle multiplayergames [14], [15]. Other works use the HLAarchitecture [16] combining classical scene graphswith simulated federations to provide interactiveJP2011-335