基于PHPA的内存大小页动态调整技术

发布时间：2019-05-20 04:12

【摘要】：为了提高物理资源的利用率,虚拟化技术得以普遍运用。通过运用虚拟化技术,单台服务器上可以整合许多虚拟机来为用户提供各种服务。为了保证用户的良好体验,这台服务器对硬件的要求就会比较高,其中也包括对内存的需求。在现有市场上的服务器,有通过基于hugetlbfs的大页管理框架来为服务应用提供2M大页,从而保证服务性能。但其不支持内存压缩,并且可扩展性较差。实验室之前的工作中,实现了一套基于PHPA(Pristine Huge Page Allocator)大页框架的的Hzmem(H代表huge,z代表压缩)系统,通过将大页内存进行压缩存放的方法,变相提高了可用的大页内存。受启发于透明大页的设计思想,在某些实际场景中,比如,服务器对大页内存的需求远远超过对小页内存的需求时,系统中可能会存在比较多的空闲小页,而缺少空闲大页。本文在实验室现有成果的基础上,设计实现了一套基于PHPA大页框架的内存大小页动态调整技术,来适应上述场景。本文的主要贡献有:1)采用PHPA大页框架,大幅度地降低了系统对内存管理的元数据开销,节省了物理内存开销。2)实现了在系统运行的时候动态调整大小页内存大小的技术。3)采用模块化的方式将本文的实现与原有的Linux内核进行了结合,不需要对原有内核进行改动。4)通过proc文件系统,实现了在用户态对内存大小页进行调整。经过功能和性能两方面的测试,本文的实现达到了预期效果,并且对原有系统的性能影响微乎其微。
[Abstract]:In order to improve the utilization of physical resources, virtualization technology has been widely used. By using virtualization technology, many virtual machines can be integrated on a single server to provide users with a variety of services. In order to ensure a good user experience, the server will be relatively high hardware requirements, including the demand for memory. In the existing market, the server provides 2m large pages for service applications through the large page management framework based on hugetlbfs, so as to ensure the performance of the service. However, it does not support memory compression, and the scalability is poor. In the previous work of the laboratory, a set of Hzmem (H representative compression (huge,z stands for compression) system based on PHPA (Pristine Huge Page Allocator) large page framework was implemented. By compression and storage of large pages of memory, the available large pages of memory were improved in disguise. Inspired by the design idea of transparent large pages, in some practical scenarios, for example, when the server's demand for large page memory far exceeds the demand for small page memory, there may be more idle pages and lack of free large pages in the system. In this paper, based on the existing laboratory results, a set of dynamic adjustment technology of memory size page based on PHPA large page framework is designed and implemented to adapt to the above scene. The main contributions of this paper are as follows: 1) the metadata overhead of memory management is greatly reduced by using PHPA large page framework. The physical memory overhead is saved. 2) the technology of dynamically adjusting the size of page memory while the system is running is realized. 3) the implementation of this paper is combined with the original Linux kernel in a modular way. There is no need to change the original kernel. 4) through the proc file system, the memory size page is adjusted in the user state. After functional and performance tests, the implementation of this paper has achieved the desired results, and has little impact on the performance of the original system.
【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】：TP316.81;TP333

【相似文献】