CAREER: Towards Interactive Simulation of Giga-Scale Agent-Based Models on Graphics Processing Units

本研究探讨大规模智能体模型(ABM)的高效仿真技术。ABMS越来越多地被用来理解许多自然、建筑和社会系统中复杂的多尺度行为。尽管ABM具有必要的结构来捕获这些系统中的复杂模型特征，但这种结构使它们在计算上具有挑战性。当前用于桌面计算的技术和对传统高性能计算的扩展不能有效地处理这种计算复杂性。这严重限制了ABM的适用性。这项研究调查了旨在利用商用图形处理单元(GPU)上可用的海量计算能力的新技术。通过有效地实现ABM仿真的超级计算，极大地扩展了基于代理的建模的可用性和适用性。此外，它还能够在廉价的台式计算机上以现实的细节水平对公共政策、救灾应急计划、药物治疗设计等方面的“假设”情景进行虚拟测试。这项研究中的主要挑战是ABM计算的重新公式，以适应GPU的数据并行模型。具体研究主题包括代理数据的表示、代理运动、复制、抽取、通信、非空间代理网络的表示和操作的函数、自适应行为、运行时用户交互、快速可视化、硬件和模型感知的自动代码优化，以及多GPU平台的多级并行。开发的技术正在应用于医学上的两个具体问题：模拟结核病和全身炎症反应综合征。这些模型能够有效地模拟疾病病理，并在电子计算机上测试新的治疗药物方案。教育主题包括课程开发，通过开发ABM主题视频游戏向K-12学生推广，本科生参与研究，以及开发一个全面的传播网页。

This research investigates techniques for efficient simulation of large scale agent-based models (ABMs). ABMs are increasingly being used to understand complex multi-scale behaviors in many natural, built and social systems. Although ABMs have the necessary structure to capture complex model characteristics in these systems, this very structure makes them computationally challenging. Current techniques for desktop computing and extensions to traditional high performance computing are incapable of efficiently handling this computational complexity. This has severely limited the applicability of ABMs. This research investigates novel techniques designed to leverage the massive computing power available on commodity graphics processing units (GPUs). It greatly expands the availability and applicability of agent-based modeling by effectively democratizing super computing for ABM simulation. Furthermore, it enables virtual testing of "what-if" scenarios in public policy, contingency planning for disaster relief, drug therapy design etc., on inexpensive desktop computers at realistic levels of detail. The main challenge in this research is the re-formulation of ABM computation tofit the data-parallel model of GPUs. Specific research topics include representation of agent data, functions for agent motion, replication, decimation, communication, representation and manipulation of non-spatial agent networks, adaptive behaviors, run-time user interaction, fast visualization, hardware and model-aware automatic code optimization, and multi-level parallelism for multi-GPU platforms. The techniques developed are being applied to two specific problems in medicine: simulation of tuberculosis and systemic inflammatory response syndrome. These models enable efficient simulation of disease pathology and in-silico testing of novel therapeutic drug protocols. Educational topics include development of courses, outreach to K-12 students through development of ABM themed video games, undergraduate involvement in research, and the development of a comprehensive dissemination web page.