Rethinking Parallel Execution, Architecture, and Software

随着处理器从单核处理器转向多核处理器，今天的计算正处于关键时刻。如果在如何编写程序以利用多个处理核心方面没有创新，软件供应商将无法再获得提高处理性能的机会，这一现象在过去40多年里一直推动着计算机行业的增长。意识到形势的严重性，许多领导人将并行计算放在了关键研究问题的首位，许多研究人员正试图建立在40年来的知识基础上，在这个关键领域取得进展。这项研究正在调查一种新的方法来解决这个关键问题，这种方法与大多数关于并行执行的主流智慧相反。该方法将广泛使用的顺序程序表示与一种新颖的动态并行执行模型相结合，以实现并行执行。这种新方法的初步经验非常有希望：它可以实现与传统技术并行化的应用程序相当的并行性，有时甚至更好，但这样做没有传统技术的缺点。这项工作调查了这种新方法在各种问题领域的适用性和有效性，包括数据库、网络和嵌入式应用程序。这项工作还将研究新的并行执行模型对处理器和芯片多处理器设计的影响。更广泛的影响包括如何编写各种重要的应用程序以在多核处理器上并行执行的潜在转变，以及由此产生的此类处理器的设计。

Computing today is at a critical juncture as processors move away from uniprocessors to multicore processors. Without innovations in how to writeprograms to make use of multiple processing cores, software vendors will no longer have access to increasing processing performance, a phenomenon that has driven the growth of the computer industry for over four decades. Realizing the enormity of the situation many leaders have put parallel computing high on their list of critical research issues and many researchers are trying to build on four decades of knowledge to make progress in this critical area. This research is investigating a novel approach to attacking this critical problem, one that is counter to much of the prevailing wisdom on parallel execution. The approach couples widely used sequential programrepresentations with a novel dynamic parallel execution model to achieve parallel execution. Initial experience with this new approach has been very promising: it can achieve parallelism comparable to, and sometimes better than, applications parallelized with traditional techniques, but can do so without the drawbacks of the traditional techniques. This work investigates the applicability and effectiveness of this new approach for a variety of problem areas, including database, networking, and embedded applications. The work will also investigate the impact of the novel parallel execution model on the design of processors and chip multiprocessors. Broader impacts include a potential transformation in how a variety of important application programs are written for parallel execution on multicore processors and the consequent design of such processors.