Collaborative Research: CISE-MSI: DP: FET: Modernizing Numerical Flow Solvers with Heterogeneous Computing

今天的计算系统通常是异质的，即配备了多核中央处理单元(CPU)和多核图形处理单元(GPU)。许多研究人员试图自行并行化他们内部的数值求解器，以同时利用多个处理器的计算能力来实现大规模模拟。然而，并行编程涉及一个陡峭的学习曲线，并且通常需要对底层硬件体系结构的深入了解才能实现最佳性能。当涉及到GPU编程时，这就成为了一个更大的问题。该项目利用了杰克逊州立大学(JSU)研究人员在基于CPU的并行计算方面的专业知识和密西西比大学(UM)研究人员在基于GPU的并行计算方面的专业知识。这个项目创建了一个通用库，其中打包了用于并行化的优化函数和例程。这个图书馆是在公共领域免费提供的。该库大大减轻了基于网格的数值求解器的并行化负担。其他研究人员可以调用这些库函数，在不同的系统上快速、轻松地并行化他们的基于网格的数值求解器，只需最少的代码更改和开发人员的努力。这项合作研究在促进研究能力的发展和在JSU和UM建立高性能科学计算方面的可持续教育方面产生了协同效应。在研究方面，该项目通过充分利用不同系统上CPU和GPU的并行计算能力来加速和优化数值流动解算器。该项目将使用OpenCL(开放计算语言)的GPU加速整合到MPI(消息传递接口)并行计算范例中。在这种组合中，MPI在多个CPU上提供粗粒度并行，而OpenCL在CPU/GPU上提供细粒度并行。这种混合以及通过共享虚拟内存实现的细粒度数据共享和同步，使数值求解器能够充分利用当今异类并行系统的计算能力。目标系统是配备了CPU-GPU异构计算节点的集群。研究重点包括MPI-OpenCL混合范式的高效通信和负载均衡。在教育方面，该项目通过在犹他州州立大学举办高性能计算(HPC)夏季学院和在密歇根大学举办HPC日活动来教授学生在现代并行计算和编程方面的必备技能，从而吸引和培养未来的劳动力。该项目由CEISE MSI计划和既定的激励竞争研究计划(EPSCoR)共同资助。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Today’s computing systems are often heterogeneous, i.e., equipped with both multi-core central processing units (CPUs) and many-core graphics processing units (GPUs). Many researchers try to parallelize their in-house numerical solvers by themselves to simultaneously utilize the computing power of multiple processors to make large scale simulation possible. However, parallel programming involves a steep learning curve and often requires the deep understanding of underlying hardware architecture to achieve optimal performance. This becomes an even bigger issue when GPU programming gets involved. This project leverages the expertise of researchers at Jackson State University (JSU) in CPU-based parallel computing and the expertise of researchers at the University of Mississippi (UM) in GPU-based parallel computing. This project creates a general-purpose library that packages optimized functions and routines used for parallelization. This library is freely available in public domain. The library greatly eases the burden of parallelization of mesh-based numerical solvers. Other researchers can call the library functions to parallelize their mesh-based numerical solvers quickly and easily on heterogeneous systems with the minimum level of code change and developers' efforts. This collaborative research generates a synergistic effect on catalyzing the development of research capabilities and establishing sustainable education in high performance scientific computing at JSU and UM.On the research side, this project accelerates and optimizes numerical flow solvers by fully exploiting the parallel computing power of both CPUs and GPUs on heterogeneous systems. This project incorporates GPU acceleration using OpenCL (Open Computing Language) into the MPI (Message Passing Interface) parallel computing paradigm. In this combination, MPI provides coarse grained parallelism on multiple CPUs and OpenCL provides fine-grained parallelism on CPUs/GPUs. This blending along with fine-grained data sharing and synchronization via shared virtual memory allows the numerical solvers to take full advantage of the computing power of today’s heterogeneous parallel systems. Targeted systems are clusters equipped with CPU-GPU heterogeneous computing nodes. Research focuses include efficient communication and load balancing on the hybrid MPI-OpenCL paradigm. On the education side, the project attracts and builds future workforce by teaching students must-have skills on modern parallel computing and programming through hosting a High Performance Computing (HPC) Summer Institute at JSU and the HPC-day event at UM.This project is jointly funded by the CISE MSI program and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.