Optimizations and Designs for Practical Improvements to Program Performance on Multicore Systems

这项研究旨在提高现代多核计算机的利用率。尽管这些现在已经很常见了，但是为这样的系统开发软件所需的技术仍然非常复杂，并且实现正确性和效率仍然是需要持续关注的问题。有两种方法可以解决这个问题。第一种方法进一步研究了一种特定形式的自动并行化：“推测多线程”。这种技术由运行时系统在内部应用，以便在没有程序员或用户干预的情况下使用多个处理器。其结果是，通过使用任何可用的多核，程序执行速度更快。这样做的好处是，结果适用于现有的遗留单线程程序，并且不需要复杂的并行编程技能，尽管性能并不总是最好的。第二种方法旨在改进对较新的编程语言和编程概念的支持，以减少现有流行的并行编程语言中存在的一些更微妙的复杂性。使用提供对低级并行编程结构的访问的编程语言可以实现良好的效率，但这也会在潜在的“数据竞争”中引入正确性问题。这类bug可以通过使用新的语言设计和习惯用法来避免，但是最好的方法仍然不清楚，并且当前的实现设计往往会导致程序性能的折衷。因此，研究工作的重点是优化具体的实现设计，并开发新的编程语言范式，以进一步降低程序员遇到的复杂性。我们的两种主要方法都旨在提高并行系统的可行性和有效性，并且通过从两个基本方向解决核心问题，我们期望对当前的实践做出重大改进。

This research aims at improving the utilization of modern, multicore computers. Although these are now commonplace, the techniques required to develop software for such systems are still very complex, and achieving both correctness and efficiency are continuing concerns. Two approaches are used to address this issue. The first furthers work on a specific form of automatic parallelization: "speculative multithreading". This technique is applied internally by the runtime system to make use of multiple processors without programmer or user intervention. The result is a program that executes faster by using any available multiple cores. This has the advantage that the result applies to existing, legacy single-threaded programs, and does not require sophisticated parallel programming skills, although performance is not always the best possible. The second approach aims at improving support for newer programming languages and programming concepts that reduce some of the more subtle complexity found in existing, popular parallel programming languages. It is possible to achieve good efficiency with programming languages that provide access to low-level parallel programming constructs, but this also introduces correctness problems in the potential for "data-races." These kinds of bugs can be prevented through the use of new language designs and idioms, but the best approach is still unclear, and current implementation designs tend to result in trade-offs with program performance. The research work thus focuses on optimizing specific implementation designs, and developing new programming language paradigms that further reduce the complexity encountered by the programmer. Both of our main approaches target improvements in feasible and efficient use of parallel systems, and by addressing the core problem from two fundamental directions we expect to make significant improvements to current practice.