SGER: Exploring Data-Parallel Techniques for Mega-Scale Agent Based Model Simulations on Graphics Processing Units

SGER：探索图形处理单元上基于大规模代理的模型仿真的数据并行技术

• 批准号: 0840666
• 负责人: Roshan D'souza
• 金额: --
• 依托单位: Michigan Technological University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2008
• 资助国家: 美国
• 起止时间: 2008-08-01 至 2010-01-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0840666&HistoricalAwards=false
• 关键词: SGER; Exploring; Data; Parallel; Techniques

This SGER award is investigating the feasibility of executing large-scale Agent-Based Models (ABMs) on Graphics Processing Units (GPUs). Agent-based modeling is a method of simulating discrete dynamic systems from the bottom up. It is a computational representation where the population behaviors result from the local interactions of the constituent entities (agents) of the system. In modeling biological systems in particular, these populations can range into the hundreds of millions. Current generation ABM simulation toolkits are designed to execute serially on the CPU. Large scale models are intractable on such systems. We are exploring newer data-parallel architectures, in particular GPUs, to execute ABMs. The challenge that we are addressing in this project is to develop novel methods for agent state representation and manipulation that align well with the computing model of the GPUs. These methods include techniques for handling collision-free motion, fast parallel memory allocation for agent replication, and efficient statistic gathering for studying asymptotic behaviors, all computed in parallel on the GPU. We are applying these methods to two problems in biomedical engineering, namely, Systemic Inflammatory Response and Hepatic Drug Disposition and comparing the results of our implementations for statistical similarity with results from implementations on CPU-based ABM toolkits such as NetLogo and SWARM. This research, if successful, will revolutionize agent-based modeling with significant impact on disciplines such as ecology, medicine, economics, and homeland security. Our techniques will effectively democratize super-computing for ABM researchers by providing tools powerful enough to simulate realistic models on inexpensive desktop computers. With the improved scalability and speed, scientists will be able to investigate entirely new emergent phenomena. Previously intractable problems such as simulating the social behavior of entire countries or simulating complete ecosystems will be possible.

该SGER奖旨在研究在图形处理单元（gpu）上执行大规模基于代理的模型（ABMs）的可行性。基于agent的建模是一种自下而上的离散动态系统仿真方法。它是一种计算表示，其中群体行为是由系统组成实体（agent）的局部相互作用产生的。特别是在对生物系统进行建模时，这些种群的数量可以达到数亿。当前一代ABM仿真工具包被设计为在CPU上串行执行。大尺度模型在这样的系统上是难以处理的。我们正在探索新的数据并行架构，特别是gpu，以执行abm。在这个项目中，我们面临的挑战是开发与gpu计算模型很好地一致的代理状态表示和操作的新方法。这些方法包括处理无碰撞运动的技术，用于代理复制的快速并行内存分配，以及用于研究渐近行为的有效统计收集，所有这些都在GPU上并行计算。我们将这些方法应用于生物医学工程中的两个问题，即系统性炎症反应和肝脏药物处置，并将我们的实现结果与基于cpu的ABM工具包（如NetLogo和SWARM）的实现结果进行统计相似性比较。这项研究如果成功，将彻底改变基于主体的建模，对生态学、医学、经济学和国土安全等学科产生重大影响。我们的技术将为ABM研究人员提供足够强大的工具，在廉价的台式计算机上模拟真实的模型，从而有效地使超级计算民主化。随着可扩展性和速度的提高，科学家们将能够研究全新的突发现象。以前棘手的问题，如模拟整个国家的社会行为或模拟完整的生态系统将成为可能。