Correct, robust and efficient concurrent computation

我的研究旨在为同时使用许多处理组件的系统所产生的问题提供正确、健壮和有效的解决方案。异步共享内存架构使用复杂的硬件和软件来加快计算速度。消息传递体系结构同样因各种消息传递服务而变得复杂，这些服务对消息排序提供了不同的约束。这些机制会导致微妙的、意想不到的行为，可能与算法设计者的预期大相径庭。因此，如果不使用同步原语，“简单”问题的“自然”解决方案是不正确的。问题是：复杂的速度增强机制使编程任务复杂化，以至于程序员增加了过多的同步，这反过来又减慢了计算速度。因此，特殊组件所带来的加速优势可能会完全失效！我研究的一个主要焦点是构建仅在这些系统带来的弱保证下是正确的算法，并且仍然利用它们的速度。我的研究的一个令人兴奋的应用是新的和新兴的多核和多核计算机。如今，即使是普通的台式电脑和笔记本电脑也有“双核”。不久的将来，“多核”架构有望在一个芯片上放置数百个不同的处理单元。科学家们对如何正确编程这些机器，同时利用它们的可用并发性的知识远远落后于他们对如何构建它们的知识。我的研究旨在定义这些机器的行为，使它们更容易编程。即使单个组件非常可靠，大型系统中的某个地方也可能出现故障，因为组件太多了。我研究的一个相关目标是设计出能够在某些组件失效的情况下继续正常运行的程序，或者在出现临时故障后自动使我们的系统恢复正常行为的程序。

My research aims for correct, robust and efficient solutions to problems that arise from systems that use many processing components simultaneously.       Asynchronous shared memory architectures use elaborate hardware and software to speed computation.  Message-passing architectures are similarly complicated by a variety of message delivery services, which provide differing constraints on message ordering.  These mechanisms cause subtle, unexpected behaviour that can differ dramatically from what algorithm designers expect.  Thus, without using synchronization primitives, "natural" solutions to "simple" problems are not correct.  The catch is: the elaborate speed-enhancing mechanisms complicate the programming task to such a degree that programmers add excessive synchronization, which, in turn, slows down computation.  Thus, the speed-up advantages intended by special components can be entirely defeated!  A major focus of my research is to construct algorithms that are correct given only the weak guarantees that come with these systems, and that still exploit their speed.       An exciting application of my research is to the new and emerging multi- and many-core computers.  Even commodity desktop and laptop computers these days have "dual cores".  The near horizon promises "many-core" architectures that place possibly hundreds of different processing units on one chip.  Scientists' knowledge of how to program these machines correctly while exploiting their available concurrency is far behind their knowledge of how to build them.  My research aims to define the behaviour of these machines so that they are easier to program.     Even when individual components are very reliable, a failure somewhere in a large system is likely because there are so many components.  A related goal of my research is to design programs that continue to operate well in spite of some dead components, or that can automatically return our systems to correct behaviour after a burst of temporary faults.