CAREER: Cooperative Redundant Threads

职业：协作冗余线程

• 批准号: 0092832
• 负责人: Eric Rotenberg
• 金额: $ 30万
• 依托单位: North Carolina State University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2001
• 资助国家: 美国
• 起止时间: 2001-03-01 至 2007-02-28
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0092832&HistoricalAwards=false
• 关键词: CAREER; Cooperative; Redundant; Threads

The proposed research enables computers to attain higher speeds, at a time when even slight performance gains are increasingly difficult to achieve by conventional means. The first key discovery is that many instructions specified by a computer program are not strictly required for producing the correct answer. That is, a much shorter program can be constructed that produces the same output in a much shorter period of time. A novel method is proposed for creating a shorter program on-the-fly by completely bypassing parts of the program predicted to be unnecessary. By skipping parts of the program, however, the processor cannot verify that it was alright to do so (because the program has been transformed). The second key discovery is that the full program can be run concurrently with the short program, in order to verify the short program is correct, without noticeably slowing down the short program. The proposed method is called the Slipstream Paradigm, after a technique in auto racing where two cars collaborate aerodynamically to speed up both cars (slipstreaming). Similarly, a slipstream processor finishes two redundant copies of the program sooner that a conventional processor finishes a single copy. Redundant execution is also transparently leveraged for fault tolerance with no additional support. Finally, the Slipstream Paradigm is implementable on emerging multithreaded processors without fundamentally reorganizing their architecture.

拟议中的研究使计算机能够获得更高的速度，而此时即使是轻微的性能增益也越来越难以通过传统手段实现。 第一个关键发现是，计算机程序指定的许多指令并不是产生正确答案的严格要求。 也就是说，可以构建一个更短的程序，在更短的时间内产生相同的输出。 提出了一种新的方法，用于通过完全绕过预测为不必要的程序的部分来在运行中创建较短的程序。 然而，通过跳过程序的一部分，处理器无法验证这样做是正确的（因为程序已经被转换）。 第二个关键发现是完整程序可以与短程序同时运行，以便验证短程序是正确的，而不会明显减慢短程序。 所提出的方法被称为滑流范式，在赛车技术后，两个汽车空气动力学合作，以加快两个汽车（滑流）。 类似地，滑流处理器完成程序的两个冗余副本比传统处理器完成单个副本更快。 冗余执行也被透明地用于容错，而无需额外的支持。 最后，滑流范式可以在新兴的多线程处理器上实现，而不需要从根本上重新组织它们的体系结构。