Virtual cloudR98922135 陳昌毅R98944033 顏昭恩R98922150 黃伯淳

2010/06/03

Outline

• Goal• Introduction– Full virtualization– Para-virtualization

• VMware• XEN• System Diagram• Current Progress• Future Work• Q&A

Goal

• Construct cloud platform that provides VM as a service

• Motivation– A Private cloud platform– Common cloud platform (Eucalyptus, ONE) doesn’t provide

migration

• Challenge– A convenient environment to migrate VM

• Access image through NFS is extremely slow

• XEN– A hypervisor– Virtualization

• Eucalyptus– Controller system– NO migration mechanism

• OpenNEbula

Related work

• We use a distributed share file system to host VM image. – VM can migrate without a central storage.

• Construct VMs as multiple prototypes according to their functionalities.

Innovations

Why Virtualization

• To increase the utilization of costly hardware resources.

• Easy to management.• Portability.

XEN

• Para-virtualization• Full virtualization (Hardware assisted)• Xend (domain-0)

System Diagram

Share File SystemPros Cons

NFS Easy to manage IO bottleneck

Glustre Distribute file system Need to modify kernel

Gluster User space distribute file system

P2P

Modified Glustre kernel is unstable to work with XEN

Migration

• Why we need live migration? We want to move VM without interrupting VM

service.Products:– Xen : live migration– VMware:VMotion

Two important consideration: 1. Downtime 2. Total migration time

Pre-copy migration

Managed Migration(Xen)

1. 1st round: – Copy all memory pages to destination machine. – Replace shadow page table to original one, and mark all pages read-

only . – Create a dirty bit map for the VM.

2. 2nd-(n-1)th round: – During the pages transferring, if VM want to modify a page, it will

invoke Xen to set the appropriate bit in the dirty bit map. – Dirty pages will be resend again.– Reset dirty bit map for the next round.

3. nth round: – When the Dirty rate is bigger than up bound, begin to do stop-and-

copy.

User Interface

User Interface

Demo- Migration

•Source

Demo- Migration

• Destination

• 13 nodes plus a master to sort 8.6GB data.• Local– All images are on local disks.

• SFS_local– All images are on local disks and are accessed via

Gluster file system.• SFS_remote– All images are on remote disks and are accessed

over Gluster file system.

Hadoop Benchmark

1 2 3 4 5 6 7 8 9 100

20406080

100120140160180

localSFS_localSFS_remote

Iteration

Performance Comparison

local SFS_local SFS_remote

87.50 103.90 160.70 Sec

Seconds

Q&A