COLLABORATIVE RESEARCH: Extreme OpenMP: A Programming Model for Productive High End Computing

拥有数百万处理器的高端分布式和分布式共享内存平台将在不久的将来部署，以解决最棘手的技术问题。它们的各个节点将是异质的多线程、多核系统，能够执行许多控制线程，但每个线程的内存相对较少，到主内存的带宽较低，并且内存层次结构较深。如果要实现这些千万亿级系统的潜力，支持在这些千万亿级系统的节点内和节点之间进行高效、可移植、高效的并行编程的编程模型是必不可少的。由于应用程序开发人员和工具供应商更容易扩展现有软件，而不是采用新的编程语言，因此基于熟悉范例的编程模型是非常理想的。OpenMP是一种得到广泛支持的共享内存编程模型，易于维护。它适合对多核节点进行编程，但不能满足分布式内存平台的需求。这项研究将极大地扩展OpenMP，使其可以用于对高端拍级系统的所有级别进行编程。为了实现这一点，调查人员将增强其现有的描述多级别并行的机制，提供用于指定同步和实现高级别局部性的附加特征，以及开发一种新的I/O接口。此外，它们将极大地提高OpenMP实现技术的最先进水平，实现节点之间以及节点内部的高性能。结果将通过最先进的Fortran/C/C++OpenMP编译器、高度优化的通信库和一系列大型应用程序进行演示。

High-end distributed and distributed shared memory platforms with millions of processors will be deployed in the near future to solve the toughest technical problems. Their individual nodes will be heterogeneous multithreading, multicore systems, capable of executing many threads of control, but with relatively little memory per thread, low bandwidth to main memory and deep memory hierarchies. A programming model that supports productive, portable, efficient parallel programming both within and across the nodes of these petascale systems is essential if their potential is to be realized. Since it is easier for application developers and tool vendors to extend existing software rather than adopt a new programming language, a programming model based upon a familiar paradigm is highly desirable. OpenMP is a widely supported shared memory programming model that provides ease of maintenance. It is suitable for programming multicore nodes, but does not address the needs of distributed memory platforms. This research will significantly extend OpenMP so that it can be used to program all levels of a high-end petascale system. In order to accomplish this, the investigators will enhance its existing mechanisms for describing multiple levels of parallelism, provide additional features for specifying synchronization and for achieving high levels of locality, as well as develop a novel I/O interface. Moreover, they will substantially improve the state of the art in OpenMP implementation technology, enabling high performance between nodes as well as within them. Results will be demonstrated via a state-of-the-art Fortran/C/C++ OpenMP compiler, a highly optimized communications library, and a range of large scale applications.