CDS&E: HAM3R: Heterogeneous Automated Management of Multiscale Methods and Resources

为了进行大规模和详细的模拟，科学家需要在不同的时间尺度上解决物理问题。将每个尺度分离为单独的模拟并将它们耦合在一起大大降低了执行时间。耦合模拟的过程是费力、复杂且容易出错的。实现性能和能源效率需要领域科学家对计算机体系结构和系统有深入的了解，并广泛地调整算法和资源请求。目前手工耦合和优化的实践状态导致了编程工作和解决方案的重复，这些工作和解决方案不能移植到新的问题和规模，以及计算机体系结构和系统中。这个跨学科的项目建立了技术，即HAM3R（多尺度方法和资源的异构自动化管理），以实现多尺度模拟的自动化耦合和优化计算和能源效率。HAM3R广泛适用于计算化学、物理、生物和材料科学等领域的多尺度问题。仿真代码的简便性和灵活性降低了领域科学家的进入门槛，对计算科学产生了广泛的影响。使领域研究人员能够利用先进的网络基础设施将加速科学吞吐量，这可能对各个学科产生变革性影响。该项目扩大了学生、代表性不足的群体和用户社区的参与。该项目开发了一个多功能软件框架（名为HAM3R），用于预测资源、工作负载、故障和电源管理，从而实现异构HPC系统上多尺度模型的性能、可扩展性和能效的自动优化。HAM3R是一个变革性的软件框架，通过支持多尺度模型的动态耦合，将API、库和运行时结合在一起，以支持广泛的耦合风格和域，从而消除了领域科学家耦合多尺度模拟的复杂性。对计算和数据瓶颈的分析产生了增强的分析和机器学习性能模型。HAM3R将为多尺度模型配备自动化资源分配，包括通过主动管理计算、通信和数据移动来预测负载平衡，从而实现异构HPC系统的可扩展性和高效模拟。本项目的智力优势推进了以下领域：(1)以数据为中心的优化，通过迁移计算和使用有损数据压缩来降低数据在尺度内和尺度间移动的成本；(2)工艺故障定制本地恢复；(3)支持智能资源管理和动态适应的尺度内和尺度间模型预测负载均衡方案；(4)跨异构设备的先进电源管理，以确保节能执行。该项目展示了HAM3R在两种不同的流行多尺度建模框架中的能力：通过域分解耦合分子动力学和晶格玻尔兹曼方法模拟；耦合耗散粒子动力学及非均质多尺度有限元模拟。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

To run large and detailed simulations, scientists need to resolve physics at different time scales. Separating each scale into individual simulations and coupling them together greatly lowers the execution time. The process of coupling the simulation is laborious, complex, and error-prone. Achieving performance and energy efficiency requires domain scientists to have deep knowledge of computer architectures and systems, and extensively tune the algorithms and resource requests. The current state-of-the-practice of manual coupling and optimization leads to duplication of programming efforts and solutions that are not portable to new problems and sizes, and to computer architectures and systems. This interdisciplinary project builds technologies, namely HAM3R (Heterogeneous Automated Management of Multiscale Methods and Resources), to automate the coupling and optimize the computing and energy efficiency for multiscale simulations. HAM3R is broadly applicable to multiscale problems in computational chemistry, physics, biology, and materials science. Greater simplicity and flexibility of simulation codes have broad impacts on computational science by reducing the entry barrier to domain scientists. Enabling domain researchers to leverage advanced cyberinfrastructure will accelerate scientific throughput, which can have transformative effects across a spectrum of disciplines. This project broadens the engagement of students and underrepresented groups and user communities. This project develops a versatile software framework (named HAM3R) for predictive resource, workload, fault, and power management that enables automated optimization of performance, scalability, and energy efficiency of multiscale models on heterogeneous HPC systems. HAM3R is a transformative software framework that removes the complexities of coupling multiscale simulations from domain scientists by enabling dynamic coupling of multiscale models that combines an API, library, and runtime to support broad coupling styles and domains. The analysis of computation and data bottlenecks yields enhanced analytical and machine-learned performance models. HAM3R will equip multiscale models with automated resource allocation -- including predictive load-balancing by proactive management of computation, communication, and data movement -- to enable scalability and efficient simulations on heterogeneous HPC systems. This project’s Intellectual Merit advances the following areas: (1) data-centric optimizations to reduce the cost of data motion intra- and inter-scale by migrating computation and using lossy data compression; (2) customized local recovery for process failures; (3) model-predictive load balancing schemes within and across scales that support intelligent resource management and dynamic adaption; and (4) advanced power management across heterogeneous devices to ensure energy-efficient execution. The project demonstrates the capabilities of HAM3R in two different popular multiscale modeling frameworks: coupled molecular dynamics and lattice Boltzmann method simulations via domain decomposition; coupled dissipative particle dynamics and finite element method simulations via heterogeneous multiscale methods.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.