Sulis: An EPSRC platform for ensemble computing delivered by HPC Midlands+

计算机模拟和建模越来越被视为现代科学的第三大支柱，仅次于理论和实验。这项活动需要越来越强大的研究计算设备。传统上，这些设施的案例是通过对更大的物理系统进行建模的科学需要，或以更高的保真度进行模拟。这种计算得益于更大、更强大的计算机，因为它在一次计算中利用了数量越来越多的计算处理单元（核心）。Sulis将支持利用并行性的替代和补充方式，特别是高吞吐量计算。这里的重点是中等规模的计算，即与那些可以在典型的高端工作站PC上在几天内执行的计算相当，但每次运行不同的输入或模型参数来解决单个问题要复制数千次。在一台多核PC上完成这种“综合”计算很容易需要几十年的时间，而在大学级别的设备上则需要几个月的时间。Sulis将允许研究人员在不到一周的时间内完成这样的工作流程，从而将他们的专业知识应用于更广泛的问题，并对新输入数据的可用性做出反应。有许多计算任务适合这种“集成计算”模型。一个相关的例子是不确定性量化（UQ）。而不是模拟一个单一的和可能不完美的模型，UQ方法产生可能的模型集合并模拟它们。这样就可以进行统计预测。模拟过程的最可能的结果可以从输出的集合中推断出来，以及基于输出的可变性的置信水平。如果使用仿真作为设计或决策工具，后者是必不可少的。类似的概念在天气预报中也很常见——模型不做绝对的预测，而是根据出现降雨的部分模拟来预测降雨的概率。这种方法适用于物理科学中的一系列问题，如预测材料特性、化学过程的产量、细菌的运动、聚变等离子体的稳定性等。其他集成计算工作流程包括优化问题。在这里，每个模拟都独立地搜索模型的输入/参数子集，从而减少了找到可行解决方案所花费的时间。这是必不可少的，例如，在研究无序材料时，比在寻求单一解决方案时假设的理想完美晶体更接近现实世界。集成计算也用于生成、采样或处理大型数据集，通常用作后续训练现代机器学习算法的输入。由于这个原因，Sulis将包括一个高容量的多拍字节的数据存储容量，利用现代固态存储技术来减少读写数据产生的瓶颈。它还将包括大量的图形处理单元（gpu）——加速器设备本身现在在机器学习应用中无处不在。对集成计算的关注给研究人员和软件工程师带来了挑战。在数千次模拟中，至少有一次失败的可能性很大。软件必须能够适应这种故障。同样，管理如此多的计算的输入和输出会使传统的数据存储子系统过载，要求用户使用计算机科学部门以外的研究人员很少遇到的数据库技术。因此，Sulis服务的一个关键特征将是研究软件工程（RSE）支持，以协助和培训用户解决这些问题，为英国研究人员应对更大规模计算挑战提供未来的竞争力。

Computer simulation and modelling is increasingly seen as the third pillar of modern science, alongside theory and experiment. Increasingly powerful research computing facilities are required for this activity. Traditionally, the case for these facilities has been made through a scientific need to model larger physical systems, or simulate with increased fidelity. Such calculations benefit from larger and more powerful computers by exploiting ever-larger numbers of computational processing units (cores) within a single calculation. Sulis will support alternative and complementary ways of exploiting parallelism, specifically high throughput computing. Here the focus is on calculations of modest size, i.e. comparable to those which could be executed on a typical high-end workstation PC in a few days, but replicated thousands of times each running with different inputs or model parameters to solve a single problem. Working through this "ensemble" of calculations could easily take decades on a single multi-core PC, or many months with university level facilities. Sulis will allow researchers to complete workflows such as this in less than a week and hence apply their expertise to a broader range of problems and be reactive to availability of new input data.There are many computational tasks which fit into this "ensemble computing" model. One pertinent example is uncertainty quantification (UQ). Rather than simulate a single and likely imperfect model, UQ approaches generate ensembles of possible models and simulates them all. This allows predictions to be made statistically. The most likely outcome of the simulated process can be inferred from the ensemble of outputs, along with a confidence level based on the variability over the outputs. The latter is essential if using simulation as a design or decision making tool. A similar concept may be familiar from weather forecasting - models do not make absolute predictions but instead predict a probability of rain based on the fraction of simulations in which this occurs. This approach is applicable to a range of problems in the physical sciences, such as predicting material properties, yield of chemical processes, the motion of bacteria, fusion plasma stability etc.Other ensemble computing workflows include optimisation problems. Here each of the simulations independently searches a subset of the inputs/parameters for a model, reducing the time taken to locate viable solutions. This is essential, for instance, in studying disordered materials, far closer to the real world than the ideal perfect crystals assumed when seeking only a single solution. Ensemble computing is also used to generate, sample or process large datasets, often for subsequent use as inputs to train modern machine learning algorithms. For this reason Sulis will include a high-capacity multi-petabyte data storage capacity, exploiting modern solid-state storage technologies to reduce bottlenecks arising from reading and writing of data. It will also include a large number of graphics processing units (GPUs) - accelerator devices themselves now ubiquitous for machine-learning applications.A focus on ensemble computing raises challenges to researchers and software engineers. With thousands of simulations, the probability that at least one will fail is substantial. Software must be resilient to this failure. Similarly, managing the input and output of so many calculations can overload traditional data storage subsystems, requiring users to work with database technology rarely encountered by researchers outside of computer science departments. Hence a key feature of the Sulis service will be Research Software Engineering (RSE) support to assist and train users in tackling these problems, future-proofing the competitiveness of UK researchers to the challenges of computing at ever larger scales.