Collaborative Research: Elements: SciMem: Enabling High Performance Multi-Scale Simulation on Big Memory Platforms

提高系统的可扩展性是提高国家科学应用计算能力的关键。然而，从内存容量的角度来看，一些应用程序经常面临可伸缩性的挑战。当处理来自不同尺度的大量模拟数据时，在多尺度模拟中尤其如此。新兴的大存储基础设施在增加模拟规模和解决更大的数值问题方面显示出巨大的潜力。然而，由于大内存机器的计算能力有限，以及大内存带来的内存异构性，使用大内存架构进行多尺度模拟具有挑战性。缺乏一个软件基础设施，可以释放大内存的全部能力来加速多尺度模拟。该项目旨在创建一种功能和软件包（名为SciMem），使大内存平台上的高性能多尺度模拟成为可能。所提出的技术为这种结构的广泛应用提供了一条途径，对科学和工程具有广泛的影响。通过学生直接参与项目和与教育活动的整合，将对他们产生影响。该项目将通过更有效地利用大型异构内存机器，在大内存平台上实现高性能多尺度模拟。具体来说，它将在异构计算系统中用预先计算和存储在内存中的数据取代计算。开发的工具SciMem将与流行的平行分子动力学模拟器LAMMPS（大规模原子/分子大规模平行模拟器）集成并进行测试。计算资源使用方面的改进将允许在新兴硬件系统上对复杂物理现象进行更精确的模型。SciMem旨在为广泛应用于计算化学和材料科学领域的某些更大规模的多尺度模拟带来10倍的性能提升，例如基于量子力学/分子力学的分子动力学（MD）催化模拟。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。

Increasing system scalability is crucial to improving nation’s computation capabilities for scientific applications. However, some applications often face the scalability challenge from the perspective of memory capacity. This is especially true in multi-scale simulations when handling massive simulation data from different scales. The emerging big memory infrastructures have shown great potential to increase the simulation scale and solve larger numerical problems. However, using big memory architectures for the multi-scale simulation is challenging, because of limited computing capability in the big memory machines and memory heterogeneity introduced by big memory. There is a lack of a software infrastructure that can release the full power of big memory to accelerate multi-scale simulation. This project aims to create a capability and a software package (named SciMem) that enables high performance multi-scale simulation on big memory platforms. The techniques presented offer a path for general use of this structure for a wide variety of applications having a broad impact on science and engineering. There will be impact on the students through their direct involvement with the project and through the integration with the educational activities.The project will enable high performance multi-scale simulations on big memory platforms through more efficient utilization of large and heterogeneous memory machines. Specifically, it will replace computations with pre-computed and stored in memory data on a heterogeneous computing systems. The developed tool, SciMem, will be integrated and tested with the popular parallel molecular dynamics simulator, LAMMPS (Large-scale Atomic/Molecular Massively Parallel Simulator). The developed improvements in the use of computational resources will allow more accurate models of complex physical phenomena to be carried out on the emerging hardware systems. SciMem aims to bring a 10x performance improvement for certain larger-scale multi-scale simulations widely applied in the fields of computational chemistry and material science, e.g., quantum mechanical/molecular mechanical-based molecular dynamics (MD) simulation of catalysis.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.