Parallel Heterogeneous Algorithms for Computational Science

高性能计算（HPC）资源正变得更容易以合理的成本获得，特别是随着可编程图形处理单元（GPU）的可用性。并行性在实践中具有挑战性，快速发展的硬件可能会导致软件维护成本增加。该提案解决了与开发动态架构感知并行科学计算应用程序相关的挑战，以应对HPC资源快速变化的性质。GPU目前包含数百个核心，当与每个CPU的核心数量的增长相结合时，会导致应用程序所需的整体并行度急剧增加。要充分利用这样的系统，必须考虑三个级别的并行性：CPU的粗粒度线程级并行性，GPU的大规模细粒度数据并行性，以及CPU/GPU集群的消息传递。这三种类型的并行性之间的平衡也需要与计算系统相匹配。因此，算法设计和实现的静态方法可能会阻碍不同类型并行计算机之间性能的可移植性。该建议侧重于新的算法方法，以找到足够的多级并行和识别和调度的任务，考虑到每个计算平台组件的性能和组件之间的通信时间。这将导致应用程序能够通过高效的调度算法和选择性能关键参数的最佳值，在运行时将任务调度到异构组件，从而适应不同的平台。工作将首先集中在两个领域：动态规划和解决方案的时间依赖性偏微分方程。在这两个领域，新的算法方法与调度异构平台相结合，将导致一系列灵活的高性能科学计算应用领域，如模拟物理和生物过程，并在生物信息学。

High performance computing (HPC) resources are becoming more readily available at a reasonable cost, particularly with the availability of programmable graphics processing units (GPUs). The parallelism is challenging to exploit in practice, and the rapidly evolving hardware can lead to increased maintenance cost of software. This proposal addresses the challenges associated with developing dynamic architecture-aware parallel scientific computing applications, to deal with the rapidly changing nature of HPC resources. GPUs presently contain hundreds of cores, which when combined with the growth of the number of cores per CPU is resulting in a dramatic increase in the overall degree of parallelism required of applications. Three levels of parallelism must be considered to fully use such systems: coarse grained thread-level parallelism for CPUs, massive fine grained data parallelism for GPUs, and message passing for clusters of CPUs/GPUs. The balance between these three types of parallelism also needs to be matched to the computing system. Therefore static approaches to algorithm design and implementation may hinder portability of performance across different types of parallel computers. This proposal focuses on new algorithmic approaches to finding sufficient multi-level parallelism and on identification and scheduling of tasks, taking into account the performance of each computing platform component and the communication time between components. This will lead to applications that are able to adapt to different platforms by scheduling tasks to the heterogeneous components at runtime through efficient scheduling algorithms and selection of optimal values of performance-critical parameters. Work will focus initially on two areas: dynamic programming and solution of time dependent partial differential equations. In both areas new algorithmic approaches combined with scheduling on heterogeneous platforms will lead to a range of flexible high performance scientific computing applications in areas such as simulation of physical and biological processes and in bioinformatics.