high performance computing on graphics processing units: hgpu.org

Recent advances in wireless/mobile communication and body worn sensors, together with ambient intelligence and seamless integrated pervasive technology have paved the way for applications operating based on social signals, i. e., sensing and processing of group behavior, interpersonal relationships, or emotions. Thinking in large, it should be apparent that modeling social systems allowing to study crowd behavior emerging from individual entities’ (agents’) condition and/or characteristics is, in fact, a challenging task. To address the heterogeneity, analytical agent-based models (ABMs) are gaining popularity due to its capability of directly representing individual entities and their interactions; unfortunately, ABMs (in which each agent has unique behavior) are not very well suited for large populations, expressed by exponentially rising simulation time. To solve this problem, the questions (i) how does the parallel execution of such models scale with capabilities of both the machine (number of cores, cluster size, etc.) and agents (behavioral adaptation function, interaction extent, etc.) and (ii) what is, in comparison, the performance coefficient applying the approach of model execution on graphical processors (GPUs) with its different pipelining architecture, need answers. To this end, we have performed simulation runs with parameter variation on a real parallel and distributed hardware platform using Cilk as well as on a GPU employing OpenCL. Simulation efficiency for two realistic models with varying complexity on a scale of 10^7 agents has shown the usefulness of both approaches.