CAREER: Interprocedural Compilers for Distributed Address Space Machines

这个项目的目标是开发新的程序间技术，用于在分布式地址空间并行机器上编译科学和工程应用程序。随着并行和超级计算体系结构的最新进展，应用程序程序员拥有前所未有的计算能力。然而，利用这种力量也变得越来越困难。并行化和优化编译器通常是不够的，因为它们将分析限制在单个过程级别。然而，通过开发一套新颖有效的程序间技术，可以显著提高并行编译器的效率，而不会显著减慢编译过程。在消息传递机器和分布式共享内存机器中，编译过程涉及几个复杂的任务，包括数据分发、位置管理、通信和/或同步的放置，以及聚合远程内存访问。该项目将开发新的程序间技术，用于在程序间执行这些分析阶段。

The goal of this project is to develop novel interprocedural techniques for compiling scientific and engineering applications on distributed address space parallel machines. With recent advances in parallel and supercomputing architectures, the application programmers have unprecedented computational power available. However, it has also become increasingly difficult to harness this power. The parallelizing and optimizing compilers usually turn out to be inadequate, since they restrict the analysis to a single procedure level. However, by developing a set of novel effective and efficient interprocedural techniques, the efficacy of the parallel compilers can be enhanced significantly, without significallty slowing the compilation process. In message passing machines and distributed shares memory machines, the compilation process involves several complex tasks, including data distribution, locality management, placement of communication and/or synchronization, and aggregating remote memory accesses. This project will develop new interprocedural techniques for performing these phases of analysis interprocedurally.