OMP-D: Shared-Address-Space Model and Programming System for High-End Computing

这项研究为分布式计算机体系结构开发了共享地址空间(SAS)编程模型。当今几乎所有的高端计算机(HEC)都具有分布式系统架构。编写此类体系结构的程序非常繁琐，并且会导致非常高的软件成本。SAS编程更容易，并带来更高的软件生产率。实现开发在分布式计算机系统上高效执行SAS程序的翻译器的既定目标可能会对软件行业和HEC的软件工程的方式产生重大影响。具体的研究是由最近成功的新的SAS模型和编译器的成功推动的，该模型和编译器已经通过成功地将OpenMP SAS模型翻译成MPI来证明执行接近手工编码的消息传递程序。该项目汇集了编程模型、并行计算编译器、编译器基础设施和HEC语言(包括HPF和OpenMP)方面的专业知识。新系统将允许当前的OpenMP程序在大型HEC平台上运行，并将允许程序员增量地移植和调试他们的顺序程序，以利用OpenMP SAS模型，同时实现比单个节点上可用的更大规模的并行性的好处。动态亲和性优化、可扩展性增强、运行时数组范围检测、多核集成和并行数据流语义等一系列新的编译器技术将被开发并在原型OMP-D(OpenMP For Distributed Architecture)翻译器中实现。性能将在NAS、SPEC OMP和SPEC MPI套件以及其他几个HEC应用程序的大量基准测试集上进行评估。

This research develops shared-address-space (SAS) programming models for distributed computer architectures. Almost all of today's high-end computers (HEC) have a distributed system architecture. Programming such architectures is tedious and contributes to very high software costs. SAS programming is easier and leads to higher software productivity. Achieving the set goals of developing translators that efficiently execute SAS programs on distributed computer systems may have a large impact on the software industry and on the way software engineering for HEC is taught.The specific research is motivated by recent successes of a new SAS model and compiler that has demonstrated to perform close to hand-coded message passing programs by successful translation of the OpenMP SAS model to MPI. This project brings together expertise in programming models, compilers for parallel computing, compiler infrastructures, and languages for HEC, including HPF and OpenMP. The new system will allow current OpenMP programs to run on large HEC platforms, and will allow programmers to incrementally port and debug their sequential programs to utilize the OpenMP SAS model while achieving the benefits of larger-scale parallelism than is available on a single node. A number of novel compiler techniques, such as dynamic affinity optimizations, scalability enhancements, runtime array range detection, multicore integration, and parallel data-flow semantics will be developed and implemented in a prototype OMP-D (OpenMP for Distributed architectures) translator. The performance will be evaluated on a large set of benchmarks from the NAS, SPEC OMP, and SPEC MPI suites as well as several other HEC applications.