1. Live Migration of Virtual Machines Christopher Clark, Keir Fraser, Steven Hand, Jacob Gorm Hansen†,Eric Jul†, Christian Limpach, Ian Pratt, Andrew Warfield University of Cambridge Computer Laboratory † Department of Computer Science,University of Copenhagen, Denmark USENIX NSDI ‘05

2. Introduction Operating system virtualization has attracted considerable interest in recent years -In data Centers, cluster computing communities allows many OS instances to run concurrently on a single physical machine Migrating an entire OS and all of its applications as one unit ◦ Compared to the process migration (residual dependencies)

3. Introduction Live Migration Without interfering the network connection Allows a separation of concerns between the users and operator of a datacenter or cluster. Allowing separation of hardware and software considerations

4. Introduction Downtime ◦ services are entirely unavailable Total migration time during which state on both machines is synchronized and which hence may affect reliability This paper use the “pre-copy” approach to achieve live migration and target on decreasing the downtime (implemented on Xen)

5. Design Network Generate an ARP reply from the migrated host, advertising that the IP has moved to a new location. Storage Use a network-attached storage (NAS) device Do not need to migrate disk storage

6. Design Memory Transfer ◦ Push phase ◦ Stop-and-copy phase ◦ Pull phase most practical solutions select one or two of the three phases ◦ pure stop-and-copy, pure demand This paper uses iterative push phase with a typically very short stop-and-copy phase.

7. Related Work Shutdown the VM Pre-Copy VMware

8. Related Work Post-Copy Live Migration of Virtual Machines Michael R. Hines, Umesh Deshpande, and Kartik Gopalan Computer Science, Binghamton University (SUNY) ACM SIGPLAN/SIGOPS VEE’09

9. Design Overview

10. WritableWorking Sets Some pages will seldom or never be modified and hence are good candidates for pre-copy Some will be written often and so should best be transferred via stop-and-copy => WritableWorking Sets

11. WritableWorking Sets

13. Dynamic Rate-Limiting Dynamically adapt the bandwidth limit during each pre-copying round The administrator selects a minimum(m) and a maximum(M) bandwidth limit The first pre-copy round transfers pages at the minimum bandwidth m

14. Dynamic Rate-Limiting Dirtying rate = ( the number of pages dirtied in the previous round) / ( duration of the previous round) Bandwidth rate for next round = Dirtying rate + 50 Mbits/sec Stop pre-copy when ◦ Calculated rate > M ◦ Less than 256KB remains to be tranferred

15. Some implementation issues Rapid Page Dirtying ◦ Do not need to always transfer hot pages Freeing Page Cache Pages ◦ In the first round Stunning Rogue Processes ◦ Limit each process to 40 write faults each time

16. Stunning Rogue Processes

17. Evaluation Dell PE-2650 server-class machines dual Xeon 2GHz CPUs 2GB memory connected via Gigabit Ethernet Storage: iSCSI protocol NAS XenLinux 2.4.27

18. a. SimpleWeb Server Apache 1.3 web server Continuously serving a single 512KB file memory allocation: 800MB Initially rate limited to 100Mbit/sec 776MB memory to be transferred in the first round 165ms outage

19. a. SimpleWeb Server

20. b.ComplexWebWorkload:SPECweb99 memory allocation: 800MB 30% require dynamic content generation 16% are HTTP POST operations 0.5% execute a CGI script The server generates access and POST logs 210ms outage

21. b.ComplexWebWorkload:SPECweb99

22. c. Low-Latency Server: Quake 3 a multiplayer on-line game server a virtual machine with 64MB of memory Six players joined the game and started to play within a shared arena transfers so little data (148KB) in the last round Downtime: 60ms

23. c. Low-Latency Server: Quake 3

24. d. A DiabolicalWorkload: MMuncher a virtual machine is writing to memory faster than can be transferred Memory: 512MB a simple C program that writes constantly to a 256MB Downtime: 3.5 seconds

25. d. A DiabolicalWorkload: MMuncher

26. Conclusion A pre-copy live migration method on Xen Concern about WWS Dynamic network-bandwidth adaption realistic server workloads such as SPECweb99 can be migrated with just 210ms downtime a Quake3 game server is migrated with an imperceptible 60ms outage