II-NEW: Enhanced Parallelization for High Performance Computing

该奖项是根据2009年美国复苏和再投资法案（公法111-5）资助的。随着科学应用程序规模的增长，现有技术提供的并行性水平导致了次优性能，因此，需要一种集成的并行性方法。本研究涉及到高性能计算集群的获取。使用这种基础设施，研究人员将提高从编译器到I/O的许多级别的并行性，以提高科学应用程序的执行性能。该基础设施提供了一个有价值的工具，可以在使用大规模科学应用的最先进系统上验证结果。本研究旨在通过创新的语言、编译器、经验调优、并行I/O和电源管理技术来提高高性能科学计算的生产力和效率。研究人员开发了流敏感静态和动态程序分析的新方法，以增强循环并行化。研究人员实现了新的推测并行化技术，为芯片多处理器（cmp）提供更高级别的线程并行性。此外，研究人员计划通过研究分布式集群的各个方面，包括新型分布式集群、复制分布式集群、异构分布式集群、自适应分布式集群和使用多个数据库的分布式集群，构建一个集成的并行I/O框架。最后，研究人员计划为并行高性能集群开发高效的能量管理方案，通过探索cmp固有的空间冗余来研究各种容错方法，并解决能量管理对系统可靠性的潜在负面影响。计算平台使研究人员能够验证他们的研究对大规模并行系统的应用程序性能和可扩展性的影响。

This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).With the growth in the size of scientific applications the level of parallelism provided by existing techniques leads to suboptimal performance and, therefore, an integrated approach to parallelism is necessary. This research involves the acquisition of a cluster for high performance computing. Using this infrastructure the investigators will improve parallelism at many levels ranging from compiler to I/O, in order to increase the execution performance of scientific applications. The infrastructure offers a valuable tool to validate the results on a state-of-the-art system using large scale scientific applications.This research aims to elevate the productivity and efficiency of high-performance scientific computing through innovative language, compiler, empirical tuning, parallel I/O, and power management technologies. The investigators develop new methods for flow-sensitive static and dynamic program analysis to enhance loop parallelization. The investigators implement new speculative parallelization techniques to expose higher levels of thread parallelism for chip-multiprocessors (CMPs). Furthermore, the investigators plan to build an integrated framework for parallel I/O by studing various aspects of declustering, including novel declustering schemes, replicated declustering, heterogeneous declustering, adaptive declustering and declustering using multiple databases. Finally, the investigators plan to develop efficient energy management schemes for parallel high-performance clusters, study various fault tolerance approaches by exploring the inherent space redundancy in CMPs, and address the potential negative effects of energy management on system reliability. The computing platform enables the investigators to validate the impact of their research on application performance and scalability on a large scale parallel system.