Development Of Advanced Computer Hardware And Software

先进计算机硬件和软件的开发

• 批准号: 10706226
• 负责人: Bernard R Brooks
• 金额: $ 100.46万
• 依托单位: NATIONAL HEART, LUNG, AND BLOOD INSTITUTE
• 依托单位国家: 美国
• 资助国家: 美国
• 起止时间: 至
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10706226
• 关键词: Adoption; Advanced Development; Affect; Algorithmic Analysis; Amber; Architecture; Carbohydrates; Cluster Analysis; Code; Common Core; Complex; Computer Hardware; Computer software; Computers; Consumption; Coupled; Crystallization; Data; Data Collection; Development; Disulfides; Evaluation; Free Energy; Future; Generations; High Performance Computing; Hydrogen; Image; Internet; Intervention; Lead; Lipid Bilayers; Lipids; Location; Manuals; Maps; Measures; Methods; Modeling; Modernization; Molecular Analysis; Molecular Conformation; Molecular Structure; Names; Noise; Oxygen; Pattern; Periodicity; Personal Computers; Plant Roots; Polysaccharides; Positioning Attribute; Potential Energy; Preparation; Process; Programming Languages; Pythons; Radial; Resources; Rotation; Sampling; Scheme; Software Tools; Specific qualifier value; Speed; Structure; Surface; System; Techniques; Technology; Testing; Time; United States National Aeronautics and Space Administration; Update; Variant; Water; Weight; Work; base; computer grid; computerized tools; cost; cost effective; data mining; density; design; disulfide bond; electronic structure; flexibility; gigabyte; improved; interest; large datasets; molecular dynamics; nanomaterials; parallel computer; parallelization; particle; programs; scientific computing; simulation; software development; theories; tool; transmission process

HL001052-25. Development of Advanced Computer Hardware and Software A Software Tool for Fast PDB-to-Parameter Generation for Molecular Dynamics Simulations We have developed a stand-alone tool called prepareforleap, implemented in the widely-used and freely-available software CPPTRAJ (which now has over 3k citations) that facilitates the preparation of structures for molecular dynamics simulations with the Amber Biomolecular simulation package. This tool parses a given PDB file and automatically handles disulfides, alternate atom locations, carbohydrates (forms, chirality, and linkages), and will change residue/atom names accordingly for use with Amber force fields. In addition to being a stand-alone program that requires no internet access, the preparation process requires no user intervention. The tool provides a curated PDB as well as the necessary commands (for e.g. bonding carbohydrates and/or creating any disulfide bonds) required to build the final system using Ambers LEaP program. We expect this tool to both help to minimize errors associated with the manual assignment of glycan parameters and considerably decrease the time-to-simulation burden. Quantifying the Effects of Lossy Compression on Energies Calculated from Molecular Dynamics Trajectories Molecular dynamics (MD) simulations can now be simulated on large systems (> 100,000 atoms) for increasingly longer time scales due to recent advances in computer software and hardware, in particular because of the adaptation of MD software to graphics processing units. Modern MD simulations often generate hundreds of gigabytes of data. As a result, there is great interest in being able to store these trajectories in as efficient a way as possible without sacrificing too much precision. We have explored how quantization and compression affects the precision of not only atomic positions (as is typically done), but also the energies calculated from such trajectories, and have compared to a wide variety of new and existing trajectory formats (21 total). We found that in particular bond energies to hydrogen are quite sensitive to precision loss, so energies calculated for systems using flexible water models (which have a large number of such bonds) require a higher precision than those using rigid water models. Based on our testing we have developed a quantization-based compression scheme using NetCDF 4/HDF5 based on the popular Amber NetCDF trajectory format that utilizes the underlying HDF5 framework to allow compression and decompression to be done on-the-fly. This new format compresses to about 66% of the size of the original NetCDF trajectory, has a positional accuracy of 5x10-5 , has an energy root-mean-square error of less than 0.1 kcal/mol. GPU-parallelization of Time-consuming Calculations in CPPTRAJ Over the past decade, graphical processing units (GPUs) have been used to increase the speed of MD simulations by several orders of magnitude. We have undertaken efforts to use GPUs not only for MD, but for the analysis of MD simulation data as well. The radial distribution function (RDF) calculation describes how density varies as a function of distance from a target particle, and is particularly important means of comparing MD simulation results to experimental results (e.g. the oxygen to oxygen RDF of a given water model). This calculation can be particularly time-consuming however since it requires evaluating a large number of distances. We have implemented a version of the RDF calculation in the freely-available analysis software CPPTRAJ on GPUs using the CUDA programming language. This enabled us to get an initial speedup of 2 orders of magnitude over the existing multi-threaded (OpenMP) CPU calculation, with some room for further optimization. In future work we plan to plan to implement more time-consuming analyses onto GPUs (such as the volumetric density map calculation). Enhancements to Cluster Analysis Calculations in CPPTRAJ Cluster analysis is a data-mining technique that can be applied to any collection of data points where a function is available to measure the distance (i.e. similarity) of those points. In the context of MD simulations, this typically means identifying important and unique conformations from trajectories that typically contain thousands to sometimes millions of structures. As such, cluster analysis is a very important tool in the analysis of MD simulations for teasing out the relevant data (and filtering the noise) from extremely large data sets. We have greatly improved the cluster analysis algorithm implemented in the freely-available analysis software CPPTRAJ in several ways. First, we have implemented the quaternion-based root-mean-square deviation (RMSD) calculation of Theobald et al., which improves the speed of calculating pairwise distances via RMSD by 10-20%. Second, we have improved the flexibility of the clustering calculation itself by allowing users to cluster using any data. This means that for example a user can cluster based on RMSD combined with energy and radius of gyration if desired, and weight each of these separately. In addition, clustering calculations can make use of user-provided distances, and clustering calculations can be restarted using previous results, or seeded with user-specified clusters. Taken together, these changes greatly expand the types of clustering analysis possible with CPPTRAJ. Autogeneration of image angles and dihedrals for crystal simulation CHARMM has been enhanced and extended to allow the autogeneration of primary-image angle and dihedral terms for all crystal types. This allows the setup and simulation of complex materials. The new code has been employed within CHARMM-GUI for setup of nanomaterials. Equilibrated lipid bilayer simulations on the GPUs We had earlier developed the Extended-eighth-shell based scheme for MPI-based parallelization of P21 periodic boundary conditions over multiple CPUs. P21 PBC is useful for the equilibration simulation of lipid bilayers as it allows the exchange of lipids between the layers during the course of simulation. In this work, we have implemented a CUDA based version of P21 PBC on the GPU. Direct space nonbonded calculations are modified via the building of the neighbor list in accordance with the half-screw rotation along the X-axis. This implementation will help in the wider adoption of the P21 PBC for lipid bilayer simulations. Multistate simulation on the GPUs We have developed a general framework to drive simulation of a reference state based on parametric weighting of using multiple states. We use a composite design pattern based scheme to handle forces from states defined by multiple PSFs and parameters. Through only python-level scripting, this method can be used for a number of interesting free-energy based methods like enveloping distribution sampling, common-core serial-atom-insertion, constant pH simulation etc. Polyrate and LAT The Polyrate software package sets a reference for the calculation of rate constants using variational transition state theory with multidimensional tunneling. The code has been rewritten to bring it to modern fortran90 standards. The code has been updated to perform least-action tunneling transmission coefficients, allowing to computing tunneling transmission coefficients. Other recent developments include new implementations of the analytical potential energy surfaces for various systems. The new implementation also includes a new interface with other programs, including to electronic structure software packages. For LoBoS hardware advancements for FY22 We have expanded our GPU pool with 20 dual A100 compute nodes featuring AMD Epyc 8 core cpus and 25G Ethernet. Along with a system feature dualing AMD Radeon 6700XT GPUs which we will use for testing code that target AMD GPUs.

HL001052-25。先进计算机硬件和软件的开发 用于分子动力学模拟的快速 PDB 到参数生成的软件工具 我们开发了一个名为prepareforleap的独立工具，在广泛使用且免费提供的软件CPPTRAJ（目前已被引用超过3000次）中实现，该工具有助于使用Amber Biomolecular模拟软件包准备用于分子动力学模拟的结构。该工具解析给定的 PDB 文件并自动处理二硫化物、替代原子位置、碳水化合物（形式、手性和连接），并将相应地更改残基/原子名称以与 Amber 力场一起使用。除了是不需要访问互联网的独立程序之外，准备过程也不需要用户干预。该工具提供了一个精心设计的 PDB 以及使用 Ambers LEaP 程序构建最终系统所需的必要命令（例如键合碳水化合物和/或创建任何二硫键）。我们希望该工具既有助于最大程度地减少与手动分配聚糖参数相关的错误，又可显着减少模拟时间负担。 量化有损压缩对分子动力学轨迹计算能量的影响 由于计算机软件和硬件的最新进展，特别是由于 MD 软件对图形处理单元的适应，分子动力学 (MD) 模拟现在可以在越来越长的时间尺度上在大型系统（> 100,000 个原子）上进行模拟。现代 MD 模拟通常会生成数百 GB 的数据。因此，人们对能够在不牺牲太多精度的情况下以尽可能有效的方式存储这些轨迹非常感兴趣。我们探索了量化和压缩如何影响原子位置的精度（通常是这样做的），以及从此类轨迹计算出的能量，并与各种新的和现有的轨迹格式（总共 21 种）进行了比较。我们发现，特别是氢的键能对精度损失非常敏感，因此使用灵活的水模型（具有大量此类键）的系统计算的能量需要比使用刚性水模型的系统更高的精度。根据我们的测试，我们使用 NetCDF 4/HDF5 开发了一种基于量化的压缩方案，该方案基于流行的 Amber NetCDF 轨迹格式，利用底层 HDF5 框架来允许即时完成压缩和解压缩。这种新格式压缩到原始 NetCDF 轨迹大小的约 66%，位置精度为 5x10-5 ，能量均方根误差小于 0.1 kcal/mol。 CPPTRAJ 中耗时计算的 GPU 并行化 在过去的十年中，图形处理单元 (GPU) 已被用于将 MD 模拟的速度提高了几个数量级。我们努力将 GPU 不仅用于 MD，还用于 MD 模拟数据的分析。径向分布函数 (RDF) 计算描述了密度如何随着距目标颗粒的距离而变化，并且是将 MD 模拟结果与实验结果（例如给定水模型的氧到氧 RDF）进行比较的特别重要的方法。然而，这种计算可能特别耗时，因为它需要评估大量距离。我们使用 CUDA 编程语言在 GPU 上的免费分析软件 CPPTRAJ 中实现了 RDF 计算的一个版本。这使我们能够比现有的多线程 (OpenMP) CPU 计算获得 2 个数量级的初始加速，并有进一步优化的空间。在未来的工作中，我们计划在 GPU 上实现更耗时的分析（例如体积密度图计算）。 CPPTRAJ 中聚类分析计算的增强 聚类分析是一种数据挖掘技术，可应用于任何数据点集合，其中可使用函数来测量这些点的距离（即相似性）。在分子动力学模拟的背景下，这通常意味着从通常包含数千甚至数百万个结构的轨迹中识别重要且独特的构象。因此，聚类分析是 MD 模拟分析中非常重要的工具，用于从极大的数据集中梳理出相关数据（并过滤噪声）。我们通过多种方式极大地改进了免费分析软件 CPPTRAJ 中实现的聚类分析算法。首先，我们实现了Theobald等人基于四元数的均方根偏差（RMSD）计算，这将通过RMSD计算成对距离的速度提高了10-20%。其次，我们允许用户使用任何数据进行聚类，从而提高了聚类计算本身的灵活性。这意味着，例如，如果需要，用户可以基于 RMSD 结合能量和回转半径进行聚类，并分别对每个进行加权。此外，聚类计算可以利用用户提供的距离，并且可以使用以前的结果重新启动聚类计算，或者使用用户指定的聚类作为种子。总而言之，这些变化极大地扩展了 CPPTRAJ 可能进行的聚类分析类型。 自动生成晶体模拟的图像角度和二面角 CHARMM 已得到增强和扩展，可以自动生成所有晶体类型的主图像角度和二面角项。 这允许复杂材料的设置和模拟。新代码已在 CHARMM-GUI 中用于纳米材料的设置。 GPU 上的平衡脂质双层模拟 我们之前开发了基于扩展八壳层的方案，用于在多个 CPU 上对 P21 周期性边界条件进行基于 MPI 的并行化。 P21 PBC 对于脂质双层的平衡模拟非常有用，因为它允许在模拟过程中在层之间交换脂质。在这项工作中，我们在 GPU 上实现了基于 CUDA 的 P21 PBC 版本。通过根据沿 X 轴的半螺旋旋转构建邻居列表来修改直接空间非键计算。这一实施将有助于更广泛地采用 P21 PBC 进行脂质双层模拟。 GPU 上的多状态模拟 我们开发了一个通用框架来驱动基于使用多个状态的参数加权的参考状态的模拟。我们使用基于复合设计模式的方案来处理由多个 PSF 和参数定义的状态的力。仅通过 Python 级脚本，该方法即可用于许多有趣的基于自由能的方法，例如包络分布采样、公共核心串行原子插入、恒定 pH 模拟等。 聚合率和 LAT Polyrate 软件包为使用多维隧道效应的变分过渡态理论计算速率常数提供了参考。该代码已被重写，以使其符合现代 fortran90 标准。该代码已更新为执行最小作用隧道传输系数，从而允许计算隧道传输系数。其他最新进展包括各种系统的分析势能面的新实现。新的实现还包括与其他程序的新接口，包括电子结构软件包。 2022 财年 LoBoS 硬件进步 我们通过 20 个配备 AMD Epyc 8 核 CPU 和 25G 以太网的双 A100 计算节点扩展了 GPU 池。连同 该系统具有双 AMD Radeon 6700XT GPU，我们将使用它来测试针对 AMD GPU 的代码。