Software: Improving Performance of Out-Of-Core Applications with Memory Access Pattern Guided Prefetching, Page Replacement and Clustering

许多大型、重要的科学和商业应用程序使用的数据集非常大，无法放入可用的RAM中。这样的应用程序不得不频繁地在主存储器和辅助存储(如磁盘)之间移动其数据集的一部分。不幸的是，由于磁盘比RAM慢几个数量级，许多大型应用程序的大部分运行时间都花在磁盘I/O上，导致性能显著下降。通常，以前的解决方案包括特定于应用程序的算法和库、基于编译器的技术以及有限的操作系统(OS)支持。虽然这些解决方案是有效的，但它们也有一些严重的缺陷。在这个方案中，PI提出了几种新方法的仔细的操作系统和体系结构联合设计，这些方法可以显著提高大型、内存密集型应用程序的性能。其基本思想是使用硬件/软件来收集和分析应用程序的运行时内存访问模式。这些模式的知识被用来指导预取和改进页面替换算法。从所提出的系统的初步结果显示，在性能方面非常有希望的改进。它们可以极大地减少页面I/O通信量，并隐藏许多页面I/O的延迟。基于NPB(NAS Parly Benchmark)基准测试的模拟结果表明，模式引导的预取和页面替换可以将性能提高100倍。所提出的方法对用户应用程序是透明的。如果这些研究成功，将对高性能计算领域产生广泛的影响。

Many large, important scientific and commercial applications use very large data sets that do not fit into available RAM. Such applications have to frequently move portions of their data sets between main memory and secondary storage (such as disks). Unfortunately, because disks are several orders of magnitude slower than RAM, many large applications spend most of their running time on disk I/Os, resulting in a significant performance degradation.A lot of research has been carried out to address this problem. Generally, the previous solutions include application-specific algorithms and libraries, compiler-based techniques, and limited Operating System (OS) support. While these solutions are effective, they also have some serious drawbacks. In this proposal, the PI proposes careful OS and architecture co-designs of several novel methods that can significantly improve the performance of large, memory-intensiveapplications. The basic idea is to use hardware/software to collect and analyze the runtime memory access patterns of applications. The knowledge of these patterns is then used to guide prefetching and to improve the page-replacement algorithms.Preliminary results from the proposed systems show very promising improvements in performance. They can drastically reduce the amount ofpage I/O traffic, and hide latencies of many page I/Os. Simulation results based on NPB (NAS Parallel Benchmarks) benchmarks show that pattern-guided prefetching and page-replacement can improve performance by up to 100 times. The proposed methods are transparent to user applications. The proposed studies, if successful, will have broad impacts on the field of high performance computing.