Exploiting GPUs to Simulate Complex Systems

This paper describes GCS (GPU-aware ComplexSim), a simulation framework for complex systems which is capable of executing on GPUs exploiting the CUDA programming model. GCS is based on an architecture similar to a previous work, ComplexSim, which provides simulation functionalities on symmetric multiprocessing (SMP) systems. With the current architecture of GCS, the simulation of the complex system run on GPUs, while tasks related to graph analysis still run on the host, by exploiting the embedded multi-thread engine of ComplexSim. The user code can be provided as a behaviour function, without taking into account issues related to task parallelisation, which are managed by the engine of GCS. In GCS, network data -" as links and mailboxes -" are organised as SoA (Structure of Array) to deal with the constraints and optimisation issues related to GPU architecture and the CUDA programming model. Moreover nodes attributes are defined by the user as in the case of ComplexSim, but GCS automatically organises them into SoA. GCS exhibits a significant improvement, in terms of simulation times, if compared to ComplexSim running on a SMP system.