SHF: AF: Large: Collaborative Research: Parallelism without Concurrency

从多核到超级计算机，并行机的广泛应用使得开发简单的编程方法变得至关重要。 在简化并行编程方面已经取得了重大进展，通过开发编程模型来支持无并发的并行性，即没有由通信进程的相对和不确定性定时引起的程序逻辑中的不确定性。 然而，实际上大多数并行程序都是并发的，因此是不确定的，导致代码只能由专家编程和理解。 本研究课题的目的是通过开发能够实现确定性并行计算的软件技术，使大多数程序员能够更容易地使用并行计算机。本研究课题从编程语言学、软件系统、算法分析和绝对性能等关键角度对该问题进行了全面的研究。 它承认并行编程不可能在每个抽象层次上都是完全确定的这一现实。 它追求三个关键策略来处理并发：封装并发，以便在适当的抽象级别上通过分层抽象隐藏它;通过重组程序来避免并发，以采用确定性方法;以及在完全封装或避免并发不切实际时管理并发。正在研究的具体技术包括交换构建块、确定性非关联还原器、确定性流水线并行、确定性接口和用于检测不变种族的广义种族检测。 该项目正在开发集成到多核软件平台中的开源库、工具和运行时扩展，以及用于比较方法的基于问题的基准测试套件。

The widespread deployment of parallel machines --- from multicores to supercomputers --- has made it critical to develop simple approaches to programming them. Significant progress has been made in simplifying parallel programming by developing programming models to support parallelism without concurrency, that is, without the nondeterminacies in the logic of programs caused by the relative and nondeterministic timing of communicating processes. Yet most parallel programs in practice are concurrent, and hence, nondeterministic, leading to code that can only be programmed and understood by experts. This research project aims to understand how parallel computers can be made easier to use by the vast majority of programmers by developing software technology that enables deterministic parallel computing.The project takes a holistic view of the problem from the key perspectives of programming linguistics, software systems, algorithmic analysis, and absolute performance. It acknowledges the reality that parallel programming cannot be fully deterministic at every level of abstraction. It is pursuing three key strategies for dealing with concurrency: encapsulating concurrency so that it is hidden by layered abstractions at appropriate abstraction levels, avoiding concurrency by restructuring programs to employ deterministic approaches, and managing concurrency when it is impractical to either encapsulate or avoid concurrency completely. Among the specific techniques being studied are commutative building blocks, deterministic nonassociative reducers, deterministic pipelined parallelism, deterministic interfaces, and generalized race detection for detecting invariant races. The project is developing open-source libraries, tools, and runtime extensions integrated into a multicore-software platform, as well as a problem-based benchmark suite to compare approaches.