Language and System Support for Petascale Irregular Applications

直到最近，大多数高性能计算应用都涉及矩阵计算。然而，新的应用领域，如计算生物学和机器学习涉及构建，计算和修改大型稀疏图。德克萨斯大学奥斯汀分校的Galois团队一直在构建一个系统，该系统利用乐观并行化和编译器分析来提高抽象级别，从而可以在不显著性能损失的情况下对这些应用程序进行编码。编程模型是顺序Java，增加了两个称为乐观集迭代器的构造。该项目汇集了Galois项目团队成员，计算生物学家和机器学习专家，以扩展和评估Galois系统在高端共享内存机器和大规模分布式内存上的应用，涉及大型稀疏图上的计算。 该项目的主要目标是（i）设计和实现适度的语言扩展，以支持嵌套数据并行，（ii）将现有系统移植到分布式内存机器，（iii）实现自适应反馈驱动的并行执行，以及（iv）开发编译器分析，以优化执行。

Until recently, most high-performance computing applications involved matrix computations. However, new application areas such as computational biology and machine learning involve constructing, computing with, and modifying large sparse graphs. The Galois team at the University of Texas, Austin has been building a system that exploits optimistic parallelization and compiler analysis to raise the level of abstraction at which these applications can be coded without a substantial performance penalty. The programming model is sequential Java augmented with two constructs called optimistic set iterators. This project brings together Galois project team members, computational biologists and machine learning experts to extend and evaluate the Galois system on high-end shared-memory machines and large-scale distributed memory for applications that involve computations on large sparse graphs. The main thrusts of the project are (i) design and implementation of modest language extensions for supporting nested data-parallelism, (ii) porting the existing system to distributed-memory machines, (iii) implementation of adaptive feedback-driven parallel execution, and (iv) development of compiler analyses to optimize execution.