SI2-SSE: Collaborative Research: An Intelligent and Adaptive Parallel CPU/GPU Co-Processing Software Library for Accelerating Reactive-Flow Simulations

为了开发下一代清洁高效的汽车发动机和发电燃烧室，工程师们需要下一代计算建模工具。准确地描述传统和替代液体运输燃料的化学成分对于预测有害排放水平和其他重要数量至关重要，但详细的化学模型的高昂计算成本对设计者的使用构成了巨大的障碍。为了使用这种精确的模型，需要能够在实际模拟中有效地处理化学问题的软件。该合作项目旨在开发此类工具，利用现代并行中央处理单元(CPU)和图形处理单元(GPU)的计算能力。除了帮助设计者创建清洁高效的发动机技术外，该项目取得的进展还广泛适用于其他计算建模问题，包括天体物理、核反应、大气化学、生化网络，甚至心脏电生理学。该项目的目标是开发专门针对GPU、CPU和其他多核加速器设备上的协同处理的软件元素，以降低使用详细化学物质的计算成本，并实现高保真但负担得起的反应流模拟。这将通过(1)开发和比较在CPU和GPU/加速器上并行运行的化学动力学积分算法，(2)开发一种检测由于化学动力学引起的局部刚度的方法，并基于现有硬件自适应地选择最有效的求解器，(3)实施计算单元聚类策略来分组相似的空间位置，(4)使用商业和开源计算流体动力学代码来模拟反应流，展示这些软件元素所提供的改进的性能，以及(5)基于上述软件元素设计可移植和可持续的软件库，包括建立用户社区。该计划的结果将是一个开放源码软件库，它将显著降低在真实燃烧模拟中使用详细、准确的化学物质的成本；该计划的成功将取决于实现数量级的性能改进或更好。

In order to develop the next generation of clean and efficient vehicle engines and power-generating combustors, engineers need the next generation of computational modeling tools. Accurately describing the chemistry of conventional and alternative liquid transportation fuels is vital to predict harmful emission levels and other important quantities, but the high computational cost of detailed models for chemistry poses a significant barrier to use by designers. In order to use such accurate models, software is needed that can efficiently handle chemistry in practical simulations. This collaborative project aims to develop such tools, employing the computational power of modern parallelized central processing units (CPUs) and graphics processing units (GPUs). In addition to helping designers create clean and efficient engine technology, the advances made in this project are widely applicable to other computational modeling problems including astrophysics, nuclear reactions, atmospheric chemistry, biochemical networks, and even cardiac electrophysiology.The objective of the proposed effort is to develop software elements specifically targeted at co-processing on GPUs, CPUs, and other many-core accelerator devices to reduce the computational cost of using detailed chemistry and enable high-fidelity yet affordable reactive-flow simulations. This will be achieved by (1) developing and comparing chemical kinetics integration algorithms for parallel operation on CPUs and GPUs/accelerators, (2) developing a method for detecting local stiffness due to chemical kinetics and adaptively selecting the most efficient solver based on available hardware, (3) implementing a computational cell clustering strategy to group similar spatial locations, (4) demonstrating the improved performance offered by these software elements using commercial and open-source computational fluid dynamics codes for modeling reactive flows, and (5) designing a portable and sustainable software library based on the above software elements, including building a community of users. The result of this program will be an open source software library that significantly decreases the cost of using detailed, accurate chemistry in realistic combustion simulations; the success of the program will be determined based on achieving order-of-magnitude performance improvement or better.