Adaptive Multi-Resolution Massively-Multicore Hybrid Dynamics

我们建议开发高度可扩展的软件，该软件将利用下一代、异类、大规模并行处理器(如广泛使用的图形处理器-GPU中的处理器)来为分子模拟中的构象采样提供数量级的性能提升。该软件将普遍适用于任何凝聚相分子体系的模拟。最初的应用领域将是在制药行业用于理性药物设计的模拟类型中加速蛋白质构象变化的采样。未来理性药物发现的应用将关键取决于模拟蛋白质构象变化和蛋白质灵活性的能力。以前计算方法在合理药物设计中的成功应用针对的是具有小的、明确定义的结合口袋的蛋白质，在蛋白质中要么相对刚性，要么在药物结合时变化很小。越来越多的药用蛋白靶标具有大的、开放的和灵活的结合位点。为了理解结合，计算模型必须能够预测这些位置在药物结合时将如何改变形状。此外，新一代药物正在开发中，这些药物针对蛋白质表面之间的相互作用，或者需要对蛋白质-蛋白质相互作用进行建模。在这些情况下，结合部位是非常动态的，因为它是在两个(或更多)结合在一起的蛋白质之间形成的。现有的分子建模算法和软件无法应对对高度灵活的蛋白质进行建模的挑战。为了确保计算科学继续在制药行业发挥重要作用，迫切需要新的软件和新的算法。我们设计了一种新的多分辨率算法，将分子动力学模拟分为两个部分：近场原子部分和远场粗粒部分。近场部分用来模拟相邻分子之间的相互作用，使用传统的原子力场，并使用标准的蒙特卡罗(MC)算法来模拟单个原子的动力学。远场部分使用粗粒(珠状)力场来模拟剩余的分子相互作用，并使用刚体动力学来模拟全球动力学(例如，大规模蛋白质构象变化)。这种动力学和分子相互作用模型的多分辨率分离，使得该算法非常适合于各种计算平台，如配备数字加速器(例如图形处理器)的超级计算机。此外，该软件还将是能源感知的，因为执行模拟的每个部分的能源成本将被考虑到决定分配哪种资源。例如，如果不立即需要模拟的结果，那么模拟可以从加速器转移，而使用低功率处理器(例如，像上网本中发现的英特尔原子集群)来运行。这将为模拟器提供最小化总模拟运行时间或总二氧化碳成本的选择。虽然该软件是为今天的集群开发的，但它将很容易扩展到未来的PETA和艾级超级计算机，在这些超级计算机中，软件适应性、能源管理和容错等概念将是实现高效扩展和高效利用超级计算机的关键。我们希望，该项目的持久影响之一将是促进国际高性能计算社区对能源感知算法和二氧化碳/能源感知调度的更好理解。我们的目的是正面解决国际高性能混凝土社区面临的问题，即增加但可变的能源成本和可用性，以及显著提高能源效率和降低高性能混凝土的环境成本的必要性。

We propose to develop highly scalable software that will exploit next generation, heterogeneous, massively parallel processors (such as those found in widely available graphics processors - GPUs) to deliver orders-of-magnitude performance increases for conformational sampling in molecular simulations. The software will be generally applicable to simulations of any condensed phase molecular system. The initial application area will be to accelerate the sampling of protein conformational change within the types of simulation used for rational drug design in the pharmaceutical industry.Future applications of rational drug discovery will depend critically on the ability to model protein conformational change and protein flexibility. Previous successful applications of computational methods in rational drug design targeted proteins that had small, well-defined binding pockets, in proteins that were either relatively rigid, or changed little upon drug binding. Increasingly, medicinally interesting protein targets have large, open and flexible binding sites. To understand binding, computational models have to be able to predict how these sites will change shape upon drug binding. Coupled to this, a new generation of drugs are being developed that target the interactions between protein surfaces, or that require modelling of protein-protein association. In these cases, the binding site is extremely dynamic, as it is formed between two (or more) proteins that have come together. Existing molecular modelling algorithms and software are incapable of stepping up to the challenge of modelling highly flexible proteins. New software and new algorithms are needed urgently to ensure that computational science continues to play an important role in the pharmaceutical industry.We have designed a new multi-resolution algorithm that will allow for the simulation of molecular dynamics to be broken into two parts; a near-field, atomistic part, and a far-field, coarse grain part. The near-field part is used to model the interactions between neighbouring molecules, using traditional atomistic forcefields, and uses a standard Monte Carlo (MC) algorithm to model the dynamics of individual atoms. The far-field part models the remaining molecular interactions using a coarse-grain (beaded) forcefield, and uses rigid-body dynamics to model global dynamics (e.g large-scale protein conformational change). This multi-resolution split of both the dynamics, and the modelling of the molecular interactions, makes the algorithm ideally suited to heterogeneous computing platforms such as supercomputers equipped with numerical accelerators (e.g. graphics processors). In addition, the software will also be energy-aware, as the energy cost of performing each part of the simulation will be factored into the decision as to which resource it is allocated. For example, if the results of the simulation were not needed immediately, then the simulation could be diverted from the accelerator, and instead run using low-power processors (e.g. clusters of Intel Atoms, like those found in netbooks). This would give the simulator the choice of minimising the total simulation runtime or the total CO2 cost. While developed for the clusters of today, the software will readily scale to the peta- and exascale supercomputers of tomorrow, where concepts such as software adaptability, energy management and fault-tolerance will be key to achieving efficient scaling and efficient supercomputer utilisation. We hope that one of the lasting impacts of this project will be a promotion of greater understanding of energy-aware algorithms and CO2/energy-aware scheduling in the international HPC community. Our intention is to tackle head-on the issues facing the international HPC community in increasing yet variable energy cost and availability, and the need to significantly improve the energy efficiency, and reduce the environmental cost of HPC.