在現今虛擬化的環境下，系統資源的分配以及效率決定該宿主機
的整合程度，然而又以記憶體較為重要，記憶體的大小決定了該宿主 機能夠容納多少虛擬機。而在眾多虛擬機之間存在著許多內容相同的 頁面，若能有有效、快速之記憶體去重複技術，則能提昇虛擬化環境 的效率。KSM(Kernel Samepage Merging)，是在 Linux 2.6.32 開始併 入主流核心當中的一項記憶體去重複化技術，該技術用來找出在宿主 機上所有虛擬機的重複頁面，並進行合併和回收。雖然 KSM 能有效地 掃描並合併重複頁面，但其執行時間仍有改進的空間，若要加快其速 度則必須付出相當程度的系統資源消耗(即CPU使用率)。故本論文提出 記憶體熱區(hot-zone)的概念並以加以應用在 KSM 的啟動階段中，藉由 找出哪些記憶體位址為最有可能合併的區域並定義為熱區，將那些被 定義為熱區的頁面提前進行掃描和合併，藉此加快剛開啟 KSM 至完成 掃描之執行時間。此外，還能比原 KSM 較早利用那些被回收之記憶體 頁面，提升系統效率。我們透過一些實驗證實 HZ KSM 較原 KSM 合併 速度更快，且消耗較少的系統資源。

In virtualized environments, memory overcommitment can enhance the machine utilization, because the number of virtual machines (VMs) that can be launched on a single physical machine simultaneously is primarily determined by the size of physical memory. One technique to achieve memory overcommitment is memory de-duplication, which detects and merges the identical memory pages across different virtual machines. The state of art implementation of memory de-duplication is vanilla KSM (Kernel Same-paging Merging), which however is slow in the start-up time, the time from activating to converging. In this thesis, we present the HZ (Hot-Zone) KSM to accelerate the start-up time of KSM. A memory page is called hot if it is likely to be merged. We have observed that different virtual machines that run the same operating systems and applications could have similar hot pages. Therefore, the memory hot zones can be learned from historical statistics. HZ KSM searches the merge-able pages in the hot-zones more frequently to shorten the start-up time. Since the number of hot pages is relatively small, HZ KSM does not increase the CPU time significantly. Experimental results show that HZ KSM can reduce the 30% of start-up time comparing to vanilla KSM.

Chinese Abstract i
Abstract ii
Contents v
List of Figures vii
List of Tables viii
1 Introduction 1
2 Background 3
2.1 Memory overcommit ........................... 3
2.1.1 Swapping ............................. 3
2.1.2 Ballooning............................. 3
2.1.3 Memory page sharing....................... 4
2.2 KSM.................................... 5
2.3 Related work ............................... 7
3 Design and Implementation 10
3.1 Overview.................................. 10
3.2 Address translation............................ 10
3.3 Memory hot-zone ............................. 11
3.4 Scanning rule ............................... 13
4 Experiments 15
4.1 Environmental setup ........................... 15
4.2 The influence of GFN list size on CPU overhead . . . . . . . . . . . . 16
4.3 Page sharing in different workloads ................... 17
4.3.1 The results of 4VMs....................... 17
4.3.2 The results of 8VMs....................... 20
4.4 CPU overhead............................... 22
4.4.1 The same basis for comparing overhead. . . . . . . . . . . . . 22
4.4.2 Average CPU utilization..................... 24
4.5 Mix workload ............................... 25
5 Conclusion 27

[1] Michael Armbrust, Armando Fox, Rean Griffith, Anthony D Joseph, Randy H Katz, Andrew Konwinski, Gunho Lee, David A Patterson, Ariel Rabkin, Ion Stoica, et al. Above the clouds: a berkeley view of cloud computing, 2009. University of California at Berkeley, 2009.
[2] Sean Kenneth Barker, Timothy Wood, Prashant J Shenoy, and Ramesh K Sitaraman. An empirical study of memory sharing in virtual machines. In USENIX Annual Technical Conference, pages 273–284, 2012.
[3] Andrea Arcangeli, Izik Eidus, and Chris Wright. Increasing memory density by using ksm. In Proceedings of the linux symposium, pages 19–28. Citeseer, 2009.
[4] Shashank Rachamalla, Debadatta Mishra, and Purushottam Kulkarni. Share- o-meter: An empirical analysis of ksm based memory sharing in virtualized systems. In High Performance Computing (HiPC), 2013 20th International Conference on, pages 59–68. IEEE, 2013.
[5] Avi Kivity, Yaniv Kamay, Dor Laor, Uri Lublin, and Anthony Liguori. kvm: the linux virtual machine monitor. page 225, 2007.
[6] Fabrice Bellard. Qemu, a fast and portable dynamic translator. In USENIX Annual Technical Conference, FREENIX Track, pages 41–46, 2005.
[7] L Adam. Manage resources on overcommitted kvm hosts, 2011.
[8] Edouard Bugnion, Scott Devine, Kinshuk Govil, and Mendel Rosenblum. Disco: Running commodity operating systems on scalable multiprocessors. ACM Transactions on Computer Systems (TOCS), 15(4):412–447, 1997.
[9] Carl A Waldspurger. Memory resource management in vmware esx server. ACM SIGOPS Operating Systems Review, 36(SI):181–194, 2002.
[10] Konrad Miller, Fabian Franz, Marc Rittinghaus, Marius Hillenbrand, and Frank Bellosa. Xlh: More effective memory deduplication scanners through cross-layer hints. In USENIX Annual Technical Conference, pages 279–290, 2013.
[11] Licheng Chen, Zhipeng Wei, Zehan Cui, Mingyu Chen, Haiyang Pan, and Yun- gang Bao. Cmd: classification-based memory deduplication through page ac- cess characteristics. In ACM SIGPLAN Notices, volume 49, pages 65–76. ACM, 2014.
[12] Yan Deng, Chunming Hu, Tianyu Wo, Bo Li, and Lei Cui. A memory dedu- plication approach based on group in virtualized environments. In Service Ori- ented System Engineering (SOSE), 2013 IEEE 7th International Symposium on, pages 367–372. IEEE, 2013.
[13] T Veni and S Bhanu. Mdedup++: Exploiting temporal and spatial page- sharing behaviors for memory deduplication enhancement. The Computer Jour- nal, 59(3):353–370, 2016.
[14] Yeji Nam, Dongwoo Lee, and Young Ik Eom. Self: Improving the memory- sharing opportunity using virtual-machine self-hints in virtualized systems. In Proceedings of the 6th Asia-Pacific Workshop on Systems, page 10. ACM, 2015.