1. Optimizing the Migration of Virtual Computers
Constantine P. Sapuntzakis, Ramesh Chandra, Ben Pfaff, Jim Chow, Monica S. Lam, and
Mendel Rosenblum
Presented by: Seth Darden

2. Outline Introduction Usages and Requirements Optimizations
Experimental Results
Summary & Conclusion

3. Introduction
Computer environments are hard to maintain and are hard to move between machines.
Capsule for a machine Architecture is defined as the data type encapsulating the complete state of a machine.
Goals :
to explore the design of a capsule-based system architecture, “The Collective”
examine its potential to provide user mobility, recovery, and simpler system management.

4. Storing and Migrating Capsules
Methods used to increase transfer of capsules and reduce the startup times:
1. Save all the updates made to disk on a separate disk, using copy-on-Write (COW).
2. Reduce the memory state of the machine by flushing non-essential data to disk.
3. Disk pages are fetched on demand as the capsule runs, taking full advantage of the operating system’s ability to tolerate disk fetch latencies.
4. Collision-resistant hashes are used to avoid sending
pages of memory or disk data that already exist at the destination.

5. Usages and Requirements
The Collective system uses serialization and mobility of capsules to bestow:
User Mobility
Backup
Software Management
Hardware Management.

6. Usages and Requirements (cont.)
User Mobility
Capsules can be taken anywhere by the user, since it is not attached to a specific machine.
Most of the state of a user’s active capsule is already present at both home and the other location (e.g. at work, lab, classroom)
Only the differences in state has to be transferred during migration
User has the ability to commence the capsule migration anytime desired.
Drawback: The system may not work well with data-intensive applications (e.g. animation development or video editing)

7. Usages and Requirements (cont.)
Backups
A user has the ability to save snapshots of their capsules as backups.
System Management
Capsules can decrease the load of having to manage software and hardware.
By creating an inactive capsule, it will be possible to distribute to multiple machines (similar to a disk image)
Drawback: moving the first capsule to a machine over the network can be costly.

8. Optimization
The optimization methods were designed to exploit the snapshot concept and the machines in a Collective system.
Ballooning
Ballooning minimizes the size of the compressed memory state
Reduces the start-up time of capsules.
Capsule Hierarchies
Capsules in the Collective system are rarely created from scratch, but are mostly derived from other capsules.

9. Balloon is filled with “0”’s
Ballooning
balloon
balloon
balloon
Inflate
deflate
processes
Balloon is filled with “0”’s

10. Capsule Hierarchies An example capsule hierarchy. VT Capsule Disk
Math Dept.
Latest Disk
CS Dept.
Latest Disk
Seth’s capsule snapshot
Latest Disk
Mario’s capsule
Latest Disk
Grace’s capsule Latest Disk
Math Dept.
Updated capsule
Latest Disk
Student2 capsule
Latest Disk
Lee’s capsule
Latest Disk

11. Capsule Hierarchies (cont.)
A capsule disk and the chain of COW disks that comprise it.

12. Demand Paging of Capsule Disks
Capsule disk are read page-by-page on demand, rather than being pre-fetched.
Starting a capsule on a machine is done as follows:
1. The memory image and all the bitmaps of the COW disks are transferred if they are not available locally.
2. The capsule is extended with a new, local latest COW disk.
3. For each remote COW disk, the corresponding
shadow COW disk is created if it does not already exist.

13. Hash Cache Protocol Typical HCP session (a) Session initiation
(b) Hash cache hit
(c) Hash cache miss

14. Experimental Results Configuration system based on:
System: VMware GSX Server 2.0.1
OS: Red Hat Linux 7.3 (kernel )
Our Experiment Configuration:
System: 2.4 GHz Pentium 4 machine
RAM: 1 GB memory
Separate computer
(to simulate a 384 kbps symmetric DSL link w/ 20ms roundtrip delay)
OS: FreeBSD
Dummynet shaper

15. Experimental Results (cont.)
Capsule Snapshots
Simulations of the migration of a user between work and home machines using a series of snapshots to determine the average completion times.
The Winstone benchmark exercises popular applications: Microsoft Office, WinZip, Norton, and Netscape
To produce the Winstone snapshots, complete images were taken of the machine state every minute during execution.

16. Capsule Snapshots gzip: Size range from 2-22 MB Transfer time: 8mins
hash + gzip:
Compresses disk to about 10-30% of raw size
hash + gzip:
Memory images range from 6-17 MB
Transfer time (DSL):
6mins

17. Summary & Conclusion
Moves a computer state over a slow DSL link in minutes rather than hours.
Capsules can be taken anywhere by the user, since it is not attached to a specific machine.
User has the ability to commence the capsule migration anytime desired.
Even though it is costly to migrate a capsule to another computer, it saves time and effort