CSR: Large: Collaborative Research: Kali: A System for Sequential Programming of Multicore Processors

今天的大多数计算机都是“多核”并行计算机，能够同时执行多个独立的计算线程。不幸的是，大多数现有的程序不是并行的，不能利用这种硬件能力。此外，使用像OpenMP这样的当前符号编写并行程序比编写顺序程序更困难，因此增加了开发成本和程序缺陷的可能性。Kali项目正在构建一个软件系统，该系统将允许大多数应用程序程序员编写顺序程序，并且在多核处理器上仍然获得良好的性能。并行性将隐藏在由专家并行程序员编写的面向对象的类库中，并将由一个复杂的运行时系统管理，该系统使用一系列根据应用程序需求定制的并行执行策略。应用程序员可以利用顺序编程的优势，如熟悉性、可读性、可维护性和可调试性。他们还将能够调整程序的性能和功率，而不必降低到较低的抽象级别。此外，Kali项目正在研究使用创新硬件，以促进高效程序的开发。最后，该项目正在产生一套应用程序基准，这将有助于类似系统的性能评估。

Most computers today are 'multicore' parallel computers that are capable of executing several independent threads of computation simultaneously. Unfortunately, most existing programs are not parallel and cannot take advantage of this hardware capability. Furthermore, writing parallel programs using current notations like OpenMP is more difficult than writing sequential programs and, as a result, increases development costs and the likelihood of program defects.The Kali project is building a software system that will permit most application programmers to write sequential programs and still obtain good performance on multicore processors. Parallelism will be hidden within object-oriented class libraries written by expert parallel programmers and it will be managed by a sophisticated runtime system that uses a range of parallel execution strategies customized to the needs of the application. Applications programmers can take advantage of the benefits of sequential programming such as familiarity, readability, maintainability, and debuggability. They will also be able to tune program performance and power without having to drop down to a lower abstraction level. In addition, the Kali project is studying the use of innovative hardware to facilitate the development of efficient programs. Finally, the project is producing a suite of application benchmarks that will be useful for performance evaluation of similar systems.