1 / 34

Difference Engine: Harnessing Memory Redundancy in Virtual Machines

Difference Engine: Harnessing Memory Redundancy in Virtual Machines. Diwaker Gupta et al., OSDI 08. presenter : Dongwoo Kang rediori@dankook.ac.kr. Contents. Introduction DE(Difference Engine) Sharing Patching Compress Implementation Evaluation Conclusion. introduction.

laban
Télécharger la présentation

Difference Engine: Harnessing Memory Redundancy in Virtual Machines

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Difference Engine:Harnessing Memory Redundancy in Virtual Machines Diwaker Gupta et al., OSDI 08 presenter : Dongwoo Kang rediori@dankook.ac.kr

  2. Contents • Introduction • DE(Difference Engine) • Sharing • Patching • Compress • Implementation • Evaluation • Conclusion

  3. introduction

  4. Virtualization • Virtualization has improved and spread over the past decade • benefit of Virtualization • High CPU utilization • Individual server machines often run at 5-10% CPU utilization • Over-provision for peak levels of demand • Isolation between services source : http://www.bcgsystems.com/blog/bid/38584/Server-Virtualization-Enabling-Technology-for-the-Masses

  5. CPU • Many core support virtualization • Can run many VMs source : http://gigglehd.com/zbxe/4164557 source : http://software.intel.com/en-us/articles/planning-for-parallel-optimization/

  6. Memory for Virtualization • 1 GB of RAM • Multiplexing 10 VMs on that same host, • however, would allocate each just 100MB • Increase memory capacity • Need Extra slots on the mother board • Need higher-capacity modules • Consume significant power • High density memory is expensive CPU’s suited to multiplexing, but main memory is not source : http://www.legitreviews.com/article/1882/4/

  7. Techniques to decrease the memory • VMware ESX server • Content-based page sharing • Reduce 10-40% on Homogeneous VMs • Disco • TPS(transparent page sharing) • But, modify Guest VM YeoungUng Park PPT in 2011 source : http://www.rpgminiatures.com/acatalog/4238_Chimera_-_War_Drums_-_Rare.html

  8. Difference Engine

  9. Overview of Difference Engine Similar Page Identical Page Difference Engine sharing patch compress source : http://www.khulsey.com/tm_ferrari_f1_engine.html

  10. 1. DE – Sharing • Scan memory to find identical pages • Whole page Sharing • How to? • Content based page sharing • Use SuperFastHashfunction • perform byte-by-byte comparison • Write Protect • Mark shared page as read-only • Writing to the shared page causes page fault trapped by VMM • VMM creates a private copy and updates virtual memory

  11. 2. DE-Patch • Patching • Sub-page sharing • Reducing the memory required to store similar pages • Create patch • Use Xdelta, Binary diff program • patch + reference page = target page • Don’t’ create patch unless it less than half a page

  12. 2. DE-Patch • Effectiveness of sub-page sharing • Snapshot based test • Each VM have 512MB • 77% Savings RUBiS Kernel Compile lmbench, Vim 7.0 Compile MIXED-1

  13. 3. DE-Compress • Compress pages that are not similar to anything • LZO, WKdm algorithms • How to select pages for compress? • Clock • NRU(Not Recently Used) Algorithm

  14. 4. DE-Paging • Good candidate page for swapping out • It would likely not be accessed in the near future • It is the same requirement as compressed/patched pages • compressed and patched pages candidate for swapping out

  15. Implementation

  16. Implementation • Implementation on the top of Xen 3.0.4 • Roughly 14,500 lines of code • Additional 20,000 lines from ports of existing algorithms • Xdelta, LZO, WKdm

  17. 1. DE-Sharing • SuperFastHash • Hash table size is limited by Xen’s Heap • Hash Table can hold entries for only 1/5 of physical memory • 1.76MB Hash Table • So, 5 passes to scan whole memory source : http://www.azillionmonkeys.com/qed/hash.html

  18. 2. DE-Patch • Identifying candidate reference pages • Uses hash of 64Byte blocks of each page ( random portion ) • HashSimilarityDetector(k,s) • k is group • s is Hash key Index : AB HashSimilarityDetector(1,1) Index : ABCD HashSimilarityDetector(1,2) HashSimilarityDetector(2,1) Index : AB Index : CD Higher values of s capture local similarity, while higher k values in corporate global similarity

  19. 2. DE-Patch • HashSimilarityDetector(k,s), c • c is number of candidates Apache RUBiS Sysbench Kernel Compile IOZone, dbench lmbench, Vim 7.0 Compile MIXED-1 MIXED-2

  20. Clock • Not-Recently Used policy • Use R/M bits • Checks if page has been referenced/modified • C1 - Recently Modified [M,R=1,1] • C2 - Not Recently Modified [M,R=1,0] • C3 - Not Recently Accessed [M,R=0,0] • C4 - Not Accessed for an extended period C1 : Ignore C2 : sharing, reference pages for patching, but cannot be patched or compressed themselves C3 : shared, patched C4 : compression, swapping

  21. Clock Performance • Lifetime of patched and compressed pages

  22. Compress • Operates similarly to patching • patch + reference page = target page • Uncompress ‘compressed page’ = target page • Interaction between compression and patching • Compress a page, the page can no longer be used as a reference for a later patched page • Sharing -> Patching -> Compress

  23. Paging • Disk paging done by swapd in Dom0 memory_monitor(){ if( mem > HIGH_WATERMARK ) then while( mem >= LOW_WATERMARK ){ then swapout ( page ) endif } endif }

  24. performance evaluation

  25. System • Dell PowerEdge 1950 • 2 processors • 2.3GHz Intel Xeon Processors • VMware ESX Server 3.0.1 • But, Use a single CPU

  26. Evaluation 1 • Homogeneous VMs • 4 VMs, each run dbench • 512MB memory • OS : Debian 3.1

  27. Evaluation 2 • Heterogeneous VMs • MIXED-1, MIXED-2

  28. Evaluation 3 • Overhead of DE • MIXED-1 Env • baseline is Xen without DE • Overhead is 7% • Overhead of VMware ESX Server is 5%

  29. Evaluation 4 • Benefits from memory saved • Run more VMs • Increase throughput

  30. Conclusion

  31. Conclusion • Main memory is a bottlenecks on Virtualization • Present Difference Engine • Whole page sharing • Page patching( sub-page sharing ) • Page compress • 1.6 ~ 2.5 more memory savings than VMware ESX Server • Saved memory can be used to run more VMs => Improve throughput

  32. Q & A

  33. Backup Slides

  34. Page Table Entry

More Related