本文選題：Linux + 內(nèi)核�。� 參考：《浙江大學(xué)》2017年碩士論文

【摘要】：在當今的云計算環(huán)境下,為了提高物理資源的使用效率,單臺物理服務(wù)器上整合了幾十甚至上百個虛擬機來承載云業(yè)務(wù)。為了保證云業(yè)務(wù)的服務(wù)質(zhì)量,物理服務(wù)器采用無交換分區(qū)的2M大內(nèi)存頁機制管理物理內(nèi)存。然而Linux現(xiàn)有的大頁管理機制Hugetlbfs由于不支持內(nèi)存壓縮回收以及大頁空間拓展性差,導(dǎo)致其不能很好地滿足業(yè)務(wù)需求。同時,在實際的部署環(huán)境中,物理服務(wù)器內(nèi)存中存在較多的重復(fù)數(shù)據(jù),而現(xiàn)有的內(nèi)存去重技術(shù)(KSM、UKSM)并不能很好地支持大頁機制,致使寶貴的物理內(nèi)存資源沒有被充分的利用。為了提高云計算環(huán)境下物理服務(wù)器系統(tǒng)的性能和物理內(nèi)存的使用效率(復(fù)用率),同時盡可能小地修改系統(tǒng)內(nèi)核。本文將基于大頁內(nèi)存管理分配框架PHPA(Pristine Huge Page Allocator),設(shè)計和實現(xiàn)了一套大頁內(nèi)存壓縮管理系統(tǒng)M2ram(2M ram)。該系統(tǒng)將不經(jīng)常使用的大頁壓縮存放在內(nèi)存中,以提高物理內(nèi)存的復(fù)用率。為了實現(xiàn)本系統(tǒng),本文完成了下列工作:(1)采用全新大頁內(nèi)存管理框架PHPA。該框架和Linux內(nèi)核耦合性低,還可以大幅降低大頁元數(shù)據(jù)的內(nèi)存開銷,進一步節(jié)省可用物理內(nèi)存。(2)對業(yè)務(wù)場景進行分析,并根據(jù)其設(shè)計一套全新的從冷熱頁追蹤、大頁回收、內(nèi)存壓縮管理到缺頁異常處理完整流程的大頁壓縮系統(tǒng)M2ram。利用多流壓縮技術(shù),提高了在NUMA架構(gòu)下M2ram的并發(fā)能力。(3)提出了把2M大頁數(shù)據(jù)壓縮后存放到4k頁空間的全新壓縮數(shù)據(jù)存儲管理機制。其空間浪費率低達1/512,不會產(chǎn)生碎片化問題。(4)相比于Hugetlbfs的串行缺頁異常處理和復(fù)雜的SwapCache機制,本文實現(xiàn)了并行化大頁缺頁異常處理和去SwapCache機制,保證了內(nèi)核的高響應(yīng)。經(jīng)仿真實驗和標準化測試,此壓縮系統(tǒng)使得內(nèi)存復(fù)用率可達到2倍以上。此系統(tǒng)的高內(nèi)存復(fù)用率、高響應(yīng)以及高魯棒性使得其有良好的工業(yè)應(yīng)用前景。
[Abstract]:In today's cloud computing environment, in order to improve the efficiency of the use of physical resources, a single physical server integrates dozens or even hundreds of virtual machines to host cloud services. In order to ensure the quality of service of cloud service, physical server manages physical memory with 2 M large memory page mechanism without switch partition. However, the existing large page management mechanism of Linux, Hugetlbfs, can not meet the business requirements due to its lack of support for memory compression and recycling and poor expansion of large page space. At the same time, in the actual deployment environment, there is more repeated data in the memory of physical server. However, the existing memory de-reloading technology (KSM / UKSMSM) can not support the large page mechanism very well, resulting in the precious physical memory resources are not fully utilized. In order to improve the performance of physical server system and the efficiency of using physical memory in cloud computing environment, the system kernel is modified as small as possible. In this paper, we design and implement a large page memory compression management system based on the large page memory allocation framework PHPA(Pristine Huge Page Allocator. In order to improve the reuse rate of physical memory, the system compresses and stores the infrequently used large pages in memory. In order to achieve this system, this paper completed the following work: 1) using a new large page memory management framework PHPAA. The low coupling between the framework and the Linux kernel can also significantly reduce the memory overhead of large page metadata, further save the available physical memory. 2) analyze the business scenarios, and design a new set of hot and cold page tracking, large page recycling, according to the framework. Memory compression management to the page exception handling the full process of the large page compression system M 2 ram. By using multi-stream compression technology, the concurrency capability of M2ram under NUMA architecture is improved. (3) A new compressed data storage management mechanism is proposed, in which 2 M large page data is compressed and stored into 4k page space. Compared with Hugetlbfs's serial page missing exception handling and complex SwapCache mechanism, this paper implements parallel large page missing exception handling and SwapCache removal mechanism, which ensures high kernel response. Through simulation and standardized test, the compression system can make the memory reuse rate more than 2 times. Because of its high memory reuse rate, high response and high robustness, the system has a good prospect of industrial application.
【學(xué)位授予單位】：浙江大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP393.09;TP311.52

【相似文獻】

相關(guān)期刊論文 前10條

1 莫建麟;王玉晶;吳U，

本文編號：1948493