CAREER: Enhanced Software Distributed Shared Memory as a Compiler Target

本研究的目的是提高易用性和性能的并行系统相结合的最先进的编译器技术与增强实现的软件分布式共享内存（DSM）。将开发实现这些目标所必需的编程接口、编译时分析和运行时支持. 以分布式系统上的底层消息传递为目标的编译器为访问模式有规律的应用程序提供了良好的性能。 然而，其他类型的访问，如果完全可编译的话，目前需要复杂的运行时支持。运行时DSM系统通过按需获取数据提供了良好的整体性能，但常规访问会产生额外的开销。该基金支持的研究旨在开发DSM接口的增强功能，该接口将利用有关程序行为的静态信息，包括数据访问和一致性要求，并为共享存储器的不同硬件支持级别以及某些体系结构中的存储器的分层结构提供统一的接口和实现。 建议的工作将有助于更好地利用分布式系统的并行计算。为了进一步实现这一目标，将开发一门并行计算课程，不仅面向计算机科学的本科生，而且面向其他科学和工程学科的潜在并行计算机用户。每一个机会都将被用来强调不同系统组件之间的相互作用，以及它们对性能和编程的影响。 ***

This research aims to improve the ease-of-use and performance of parallel systems by combining state-of-the art compiler technology with an enhanced implementation of software distributed shared memory (DSM). The programming interface, compile- time analysis, and runtime support necessary to achieve these goals will be developed. Compilers that target the underlying message passing on distributed systems offer good performance for applications whose access patterns are regular. However, other types of accesses, if compilable at all, currently require complex run-time support. Run-time DSM systems provide good overall performance by fetching data on-demand, but incur additional overhead for regular accesses The research supported by this grant aims to develop enhancements to the DSM interface that will utilize static information about program behavior, including data accesses and consistency requirements, and provide a uniform interface and implementation for different levels of hardware support for shared memory, as well as for the hierarchical structure of memory in some architectures. The proposed work will contribute to the better utilization of distributed systems for parallel computing. To further this goal, a course on parallel computing will be developed that will target not only undergraduate students in computer science, but also potential parallel computer users in other science and engineering disciplines. Every opportunity will be used to stress the interaction among different system components, and their effect on performance and programming. ***