II-NEW: A Heterogeneous Testbed for Exploring Emerging HPC Tools, Programming Languages, and Applications

并行计算正处于激动人心的变化之中。几乎所有的新处理器都有多核，使所有应用程序都面临并行性。GPU是许多计算密集型应用程序的首选处理器，这使得异构性成为高性能计算环境中的另一个现实。这些特点给应用程序开发人员和系统级软件架构师带来了严峻的挑战。该项目计划开发基础设施，以支持对新兴HPC系统的编程模型、工具和运行时系统的研究。该项目创建了一个具有新兴高端系统核心方面的异类计算集群。通过将每个节点的核数最大化到最多32个，该系统将在适当数量的节点上具有高度的并行性。它利用多个异类CPU架构并部署计算加速器。该存储系统基于固态硬盘(SSD)，这是一项新兴技术，其延迟比磁性硬盘驱动器低两个数量级。对该系统的研究主要集中在执行工具链的各个方面。支持的具体研究项目包括：用于多核架构和加速器的OpenMP编译器和运行时系统；在异质环境中计算和通信操作的自动优化；在异质和志愿者环境中的高度容错执行；以及针对低延迟存储设备的并行I/O操作的优化。此外，与应用科学家的合作有望导致为各种应用领域开发高级并行化策略。

Parallel computing is in the midst of exciting changes. Virtually all new processors have multiple cores exposing all applications to parallelism. GPUs are processors of choice for many compute intensive applications, making heterogeneity another reality of the HPC landscape. These characteristics pose tough challenges to application developers as well as system level software architects. This project plans to develop the infrastructure to support research into programming models, tools and runtime systems for emerging HPC systems.The project creates a heterogeneous compute cluster with core aspects of emerging high-end systems. The system will have a high degree of parallelism distributed on a modest number of nodes by maximizing the cores per node up to 32. It utilizes multiple heterogeneous CPU architectures and deploy compute accelerators. The storage system is based on solid-state disks (SSDs), an emerging technology that allows two orders of magnitude lower latencies than magnetic hard drives. The research conducted on this system focuses on various aspects of the execution tool chain. Specific research projects supported include: OpenMP compiler and run-time system for multi-core architectures and accelerators; automatic optimization of compute and communication operations in heterogeneous environments; highly fault-tolerant execution in heterogeneous and volunteer environments; and optimization of parallel I/O operations for low latency storage devices. Furthermore, collaborative work with applications scientists is expected to lead to the development of advanced parallelization strategies for various application domains.