Sustainable domain-specific software generation tools for extremely parallel particle-based simulations

这个研究项目来自EPSRC于2010年1月举行的极限计算沙坑。沙坑过程促进了一群化学家和土木工程师之间的讨论，他们都使用基于颗粒的数值模拟方法。激发这些科学家进行模拟的应用有很大的不同（从量子力学到大坝工程），模拟的尺度非常不同（一些情况下是原子，另一些情况下是毫米大小的沙粒），粒子之间的相互作用以明显不同的方式建模。然而，所有的方法都有一个共同的基础（它们模拟粒子及其相互作用），并且它们都需要在高性能计算（HPC）环境中实现。事实上，模拟方法的相似性意味着实现有效的HPC实现存在关键的共同挑战。高性能计算最近和未来的发展使大规模并行计算成为可能；因此，在这些对工业和社会产生巨大影响的数值模型中，实现复杂性和现实性水平的阶梯式变化是有可能的。然而，与此同时，HPC硬件复杂性的增加对充分利用这些资源构成了障碍。两位计算机科学家与化学家和工程师合作，制定了应对这些挑战的策略。提出的方法是开发一种新的粒子科学语言（PSL），作为开发基于粒子模型的科学家和在快速发展的高性能计算环境中实现这些模型所需的计算机科学之间的桥梁。科学家和工程师将在这个相对简单的PSL和自动代码生成系统中实现他们的物理模型，然后自动生成优化的代码，以便在高性能系统上有效地运行模拟。研究策略是快速开发PSL的第一个实现，并在微米到毫米尺度的离散单元建模（DEM）动力学代码上进行原型设计。从这个早期的基础开始，我们将开发PSL，其关键目标是在粒子动力学领域展示应用广度的通用性，并可移植性到多种HPC架构。一系列基于粒子的方法（PBM）目前被用于模拟化学、工程、物理和生物物理学中的材料。在提议中直接考虑的四种PBM类型是分子动力学（MD），基于ONETEP量子力学的程序，离散元素建模（DEM）和光滑粒子流体动力学（SPH）。总体研究目标是开发一种可持续的工具，在未来将提供适用于从水坝工程到原子药物设计等应用的前沿研究。在目前的项目中也将取得可衡量的成就。事实上，我们预计在项目生命周期内至少有3个物理科学第一：(1)第一个描述整个蛋白质-药物复合物和溶剂的蛋白质药物模拟。(2)第一次在原子水平上使聚合物在CO2中的溶解度合理化，具有化学精度。(3)基于第一性原理量子力学的碳纳米管振动谱模拟。从计算机科学的角度来看，该项目将导致一个新的高级编译器优化框架，该框架可以捕获高度不规则数据结构背景下的依赖性、数据分布和循环巢优化，超出了多面体模型等经典方法的范围，甚至可以在形状分析和所有权类型抽象方面获得新的结果。

This research project emerged from the EPSRC Sandpit on Extreme Computing that took place in Jan 2010. The sandpit process facilitated discussion between a group of chemists and civil engineers who all use particle-based numerical simulation approaches. The applications motivating these scientists' simulations vary greatly (from quantum mechanics to dam engineering), very different scales are simulated (atoms in some cases, mm sized sand particles in others) and the interactions between the particles are modelled in significantly different ways. However all of the methods have a common basis (they simulate particles and their interactions) and they all require implementation in a high performance computing (HPC) environment. In fact the similarities in the simulation approach mean that there are key common challenges to achieving effective HPC implementations. Recent and likely future developments in HPC are making massively parallel computations viable; consequently there is a real possibility of achieving a step change in the level of complexity and realism in these numerical models with huge impact on industry and society. However, at the same time, the increased complexity of HPC hardware poses a barrier to being able to fully harness this resource. Working with the chemists and engineers, two computer scientists developed a strategy to address these challenges. The proposed approach is to develop a new Particle Science Language (PSL) to act as a bridge between the scientists who are developing particle based models and the computer science required to implement them in a rapidly evolving HPC environment. The scientists and engineers will implement their physical models in this relatively simple PSL and an automatic code generation system then automatically generate optimized code to effectively run the simulations on high performance systems. The research strategy is to rapidly develop the first implementation of the PSL and prototype it on micron to millimetre scale Discrete Element Modelling (DEM) dynamics code. From this early foundation, we will develop the PSL, with the key objectives of demonstrating generality in the breadth of applications within the particulate dynamics domain and portability to multiple HPC architectures. A range of particle based methods (PBM) are currently used to simulate materials in chemistry, engineering, physics and biophysics. The 4 types of PBM considered directly in the proposed are molecular dynamics (MD), the ONETEP quantum mechanics-based program, discrete element modelling (DEM), and smoothed particle hydrodynamics (SPH). The overall research objective is to develop a sustainable tool that will deliver, in the future, cutting edge research applicable to applications ranging from dam engineering to atomistic drug design. Measureable achievements will also be made during the current project. In fact, we anticipate at least 3 physical science firsts in the project lifetime:(1) The first protein drug simulations that describe the entire protein-drug complex and solvent from first principle quantum mechanics.(2) The first rationalization, at an atomic level, of polymer solubility in CO2, with chemical accuracy. (3) Simulation of the vibrational spectrum of a carbon nanotube from first principles quantum mechanics.(4) Simulation of triaxial compression tests with 1,000,000 particles From a Computer Science perspective the project will lead to a novel high-level compiler optimisation framework that captures dependence, data distribution and loop nest optimisation in a context of highly-irregular data structures beyond the scope of classical approaches such as the polyhedral model, or even new results in shape analysis and ownership-type abstractions.