Worked in the lab developing real-time traffic simulation using GPGPU style computation to accelerate the autonomous agent traffic computations.
Title: Large Scale Traffic Simulation on Graphics Hardware
Year: 2008
The use of virtual environments is interesting if it closely mimics real world scenarios. An interesting virtual environment is characterized by autonomous agents inhabiting the virtual world. A large scale simulation of entities in a virtual environment introduces dynamic activity into a virtual environment, closely resembling an urban area or vehicles in a freeway. However, support of large scale simulation demands high computational resources. The work presented in this thesis addresses this issue and massively expands the scale of traffic simulation.
The approach to massive expansion to the scale of traffic simulation is achieved by augmenting the Central Processing Unit with additional computational resource available from the Graphics Processing Unit or a “GPU” in a desktop. A modern Graphics Processing Units are parallel computational engines with up to 128 vector processors operating in parallel. A GPU framework is designed and implemented to support traffic simulation. A working model in a GPU follows the paradigms of stream computing. The datastructures used in GPU framework are textures units and vertex buffer objects. The shaders contain code that operates on an input stream and writes output to a predetermined initialized texture memory unit. The use of extensions such as Framebuffer Object, Vertex Buffer Object, Transform feedback, Fragment Shader, and Geometry Shader form an integral part in functioning of the GPU subsystem.
The framework built into the GPU subsystem functions in parallel with the CPU traffic simulation system. The data transfer between the CPU and the GPU via the system bus is kept to minimum to reduce latency. The GPU subsystem operates independent to its CPU built framework. The use of GPU subsystem has massively expanded the scale of traffic simulation by at least 100 times from 200 to 20,000 vehicles simulated in real time. This expansion is achieved in real time. The work demonstrates the feasibility of using a Graphics Processing Unit for computation intensive applications in real time for massive scale up in performance of applications.