XPS: FP: Collaborative Research: Parallel Irregular Programs: From High-Level Specifications to Run-time Optimizations

今天的高性能超级计算机和明天的普通计算机的核心数量都在不断增加。有效地利用这些计算机将使科学的各个领域的研究取得进展，从了解大脑到了解基本粒子再到了解宇宙。这些新计算机变得越来越复杂，编程也越来越困难。与此同时，科学和工程中使用的标准算法正在演变，越来越难以与这些机器相对应。在编程模型、工具和实现方面的进步使复杂算法的实现变得更简单，同时获得高性能，这对于使高性能计算成为所有科学家的标准工具至关重要。通常用矩阵表示的规则算法推动了高性能计算。越来越多的人对使用大规模计算机进行不规则算法很感兴趣。不规则算法出现在处理图形、稀疏矩阵、树、自适应网格等方面，并日益成为计算科学家使用的标准工具。在高级别上表达这些算法允许高性能运行时在共享内存机器上获得与这些算法的最佳手工编码实现相媲美的性能。高级描述将程序员从并行编程的复杂性中解放出来。PI正在构建运行时和编译器，以允许在分布式内存的大型计算机上执行复杂的、不规则的算法。高级表示允许系统利用关于算法语义的大量知识来优化通信、屏蔽延迟并实现高性能。

The high performance super-computers of today and the ordinary computers of tomorrow have an ever-increasing number of cores. Utilizing these computers efficiently will allow research advancements in every field of science, from understanding the brain to understanding the fundamental particles to understanding the cosmos. These new computers are increasingly complex and difficult to program. At the same time, standard algorithms used in science and engineering are evolving and are increasingly hard to map to these machines. Advances in programming models, tools, and implementations which make implementing complex algorithms simpler, while achieving high performance, are essential to making high performance computing a standard tool of all scientists.Regular algorithms, usually expressed with matrices, have driven high performance computing. Increasingly there is considerable interest in using large-scale computers for irregular algorithms. Irregular algorithms arise in manipulating graphs, sparse-matrices, trees, adaptive meshes, etc and are increasingly a standard tool used by computational scientists. Expressing such algorithms at a high-level has allowed high-performance run-times to achieve performance comparable to the best hand-coded implementations of these algorithms on shared-memory machines. A high level description frees the programmer from the complexities of parallel programming. The PIs are building run-times and compilers to allow the execution of complex, irregular algorithms on distributed-memory, large-scale computers. A high-level representation allows the system to exploit considerable knowledge about the semantics of the algorithm to optimize communication, mask latency, and achieve high-performance.