NSF Young Investigator: Concurrent Object-Oriented Programming Support for Irregular Parallel Applications

为了让并行计算成为各种计算问题的选择工具，而不仅仅是规则的数值问题，不规则并行的表达和开发必须是高效的。通过1)用并发面向对象程序设计语言支持不规则并行的表达，2)用高性能的路由网络和网络接口支持不规则并行的开发，解决了可编程性方面的主要挑战。并发面向对象语言使程序员能够隐藏并发性和分布性，允许增量并行，并将不规则并行引起的复杂性限制在程序的一小部分。然而，高效的实现需要在底层编译器和运行时系统中进行积极的局部性和粒度优化。该项目正在探索积极的类型推理和动态数据结构分析技术，以实现这样的程序优化。类型推理用于发现程序控制流，动态数据结构分析近似于程序存储图。运行时技术也在探索之中。这些技术体现在公开可用的伊利诺伊州音乐会系统中。该系统正被用于开发多个并行应用程序。

For parallel computing to become the vehicle of choice for a broad variety of computing problems, not just regular numeric ones, the expression and exploitation of irregular parallelism must be efficient. This research addresses the major challenge, programmability, by 1) supporting the expression of irregular parallelism with concurrent object-oriented programming languages and 2) supporting the exploitation of irregular parallelism with high performance routing networks and network interfaces. Concurrent object-oriented languages enable programmers to hide concurrency and distribution, allowing incremental parallelization and also confining the complexity arising from irregular parallelism to a small part of the program. However, efficient implementation requires aggressive locality and grain size optimizations in the underlying compiler and runtime system. This project is exploring aggressive type inference and dynamic data structure analysis techniques to enable such program optimizations. Type inference is used to discover program control flow, and dynamic data structure analysis approximates the program storage map. Runtime techniques are also being explored. These techniques are embodied in the publicly available Illinois Concert System. This system is being used to develop a number of parallel application programs.