一种基于物理大页的页迁移策略的设计与实现

发布时间：2018-05-18 09:08

本文选题：页迁移 + 物理大页　；参考：《浙江大学》2017年硕士论文

【摘要】：物理内存是现代计算机的核心硬件资源,同时也是短缺的资源。内存中越来越多的进程数量以及日益增长的应用程序的内存占用,都让物理内存利用率的优化永远不能停下脚步。在众多物理内存利用率的优化方案中,大页的运用是一个典型的代表。大页技术在减少TLB(TranslationLookasideBuffer,转址旁路缓存)失效开销的同时,能减少操作系统管理的页表项数量,被现代计算机系统广泛支持。目前Linux系统的大页机制采用的是基于连续小页的复合大页。在这种机制下,页描述符所描述的页框依然是小页页框;需要大页的时候,系统分配连续的小页当成一个大页使用。复合大页兼容了原有的基于小页的物理内存管理模块,而且一般能切实有效地减少系统的页表项和降低TLB的失效开销。但是,这种复合大页技术没有减少管理物理内存的元数据开销,对大页的操作常常要分解成对小页操作的迭代。在空间和时间两个方面,大页技术都有进一步优化的可能性。一种可能的优化方案是物理大页(页描述符直接描述大页)的运用,但是物理大页的运用面临着挑战。首先,一般情况下物理小页应该得到保留。大页在适用场景中确实能显著提升整机性能和节约内存空间,但是大页的适用还是不如小页广泛。如果纯用大页,可能造成过多的内部碎片,反而没有提高内存利用率。如果不放弃小页,那么物理大页就要与物理小页共存,这为操作系统的设计提出了不少问题,比如物理内存空间如何分布,如何调整大页和小页的空间配额等等。本文从一种面向Linux系统下的hugetlbfs接口的物理大页技术出发,探讨其可行的物理内存空间布局。在一种物理大页与小页共存的布局下,设计与实现了一种页迁移方案来支持大小页内存池的运行时调整。本文的主要工作有:研究分析了 Linux系统支持的大页技术,包括从系统程序员的角度看到的共享大页内存和映射大页内存,以及从内核程序员的角度看到的透明大页(THP)以及hugetlbfs,同时还研究分析了一种基于Linux的面向hugetlbfs的物理大页技术。实现了物理大页上的页迁移机制,并运用大小页的页迁移机制,设计了一种混合页粒度系统的页迁移方案。实验评测了这种页迁移方案,并对内存利用率的优化进行了量化分析。
[Abstract]:Physical memory is not only the core hardware resource of modern computer, but also the scarce resource. The increasing number of processes in memory and the memory footprint of growing applications make the optimization of physical memory utilization never stop. In many optimization schemes of physical memory utilization, the use of large pages is a typical example. Large page technology can reduce the overhead of TLBU Translation Lookaside buffer and reduce the number of page table items managed by the operating system. It is widely supported by modern computer systems. At present, the large page mechanism of Linux system adopts compound large page based on continuous small page. In this mechanism, the page frame described by the page descriptor is still a small page frame; when large pages are needed, the system allocates consecutive pages as a large page. The composite large pages are compatible with the original physical memory management module based on small pages, and generally can effectively reduce the page table items and the TLB failure overhead. However, this composite large page technique does not reduce the metadata overhead of managing physical memory, and operations on large pages are often decomposed into iterations of small page operations. In both space and time, large-page technology has the possibility of further optimization. One possible optimization scheme is the use of physical large pages (page descriptors directly describe large pages), but the use of physical large pages faces challenges. First of all, physical pages should normally be retained. Large pages do improve overall performance and save memory space in scenarios, but large pages are not as wide as small pages. If the use of large pages, may cause too much internal fragmentation, but not improve memory utilization. If we don't give up the small pages, the physical pages will coexist with the physical pages, which raises many problems for the design of the operating system, such as how to distribute the physical memory space, how to adjust the space quota of the large pages and the small pages, and so on. In this paper, a physical large page technology for hugetlbfs interface in Linux system is introduced, and its feasible physical memory space layout is discussed. Under a physical layout of large and small pages, a page migration scheme is designed and implemented to support runtime adjustment of large and small page memory pools. The main work of this paper is as follows: the large page technology supported by Linux system is studied and analyzed, including shared large page memory and mapped large page memory from the point of view of system programmer. From the point of view of kernel programmer, the paper also studies and analyzes a kind of physical large page technology for hugetlbfs based on Linux, as well as the transparent large page (THP) and hugetlbfs. The page migration mechanism on physical large pages is implemented, and a page migration scheme of mixed page granularity system is designed by using the page migration mechanism of large and large pages. The page migration scheme is evaluated and the optimization of memory utilization is analyzed quantitatively.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP333

【相似文献】