Efficient Execution of Fine-Grained Concurrent Programs

细粒度并发程序为可移植的并发编程提供了可能性，但它们的有效执行还不实用。该项目的目标是开发通过转换细粒度程序的执行粒度来支持细粒度程序高效执行的技术。细粒度程序会带来特殊的问题，因为每个粒度的开销可能会压倒利用其并发性的好处。该项目专注于并发的、基于对象的程序，因为它们允许表达细粒度的并发，并将数据和程序耦合在一起，提供重要的数据局部性信息。程序转换通过对对象进行分组来增加程序的局部性。这允许将较小的计算颗粒合并为较大的颗粒，从而提高执行效率并降低通信要求。这项研究将研究并发面向对象程序中控制和通信模式的定量分析，程序分析和转换技术的发展，以及基于可共享中间形式的实现和评估。定量分析为理解程序的行为提供了有用的参考点和基础。分析技术识别调用路径和静态对象关系，使得能够应用程序转换来沿着调用路径对对象进行分组。这样的技术增加了程序的局部性，减少了通信，并允许使用更有效的、专门的代码序列。

Fine-grained concurrent programs offer the potential for portable concurrent programming, but their efficient execution is not yet practical. The objective of this project is to develop techniques to support efficient execution of fine-grained programs by transforming their execution grain size. Fine-grained programs present special problems because the overhead of each grain can overwhelm the benefits of exploiting their concurrency. The project focuses on concurrent, object-based programs because they allow the expression of fine-grained concurrency and couple data and program, providing important data locality information. The program transformations increase program locality by grouping objects. This allows small computational grains to be merged into larger grains, increasing execution efficiency and reducing communication requirements. The research will examine quantitative analysis of control and communication patterns in concurrent object-oriented programs, the development of program analysis and transformation techniques, and their implementation and evaluation based on a sharable intermediate form. Quantitative analysis provides a useful point of reference and the basis for understanding the behavior of programs. The analysis techniques identify invocation paths and static object relations, enabling the application of program transformations grouping objects along invocation paths. Such techniques increase program locality, reducing communication and permitting the use of more efficient, specialized code sequences.