Modern Web Cloud Architecture based on Google Technologies
Grid2012 VC-Migration: Live Migration of Virtual Clusters in the Cloud
1. VC-Migration: Live Migration of
Virtual Clusters in the Cloud
Kejiang Ye, Xiaohong Jiang, Ran Ma, Fengxi Yan
CCNT Lab, College of Computer Science
Zhejiang University, China
GRID 2012
Sep. 20-23, 2012 Beijing, China
2. Outline
Motivations
VC-Migration Framework
Virtual Cluster Performance
Cross-Domain Virtual Cluster
Virtual Cluster Scalability
Dynamic Migration Strategies of Virtual Cluster
Effects of Memory Configuration
Migration Scalability
Concurrent Migration with Various Granularities
Mutual Migration
Multi-VC Migration Strategies
Live Migration of Homogeneous Virtual Clusters
Live Migration of Heterogeneous Virtual Clusters
Master Node Migration vs. Slave Node Migration
Related Work
Conclusion & Future Work
GRID 2012
Sep. 20-23, 2012 Beijing, China
3. Outline
Motivations
VC-Migration Framework
Virtual Cluster Performance
Cross-Domain Virtual Cluster
Virtual Cluster Scalability
Dynamic Migration Strategies of Virtual Cluster
Effects of Memory Configuration
Migration Scalability
Concurrent Migration with Various Granularities
Mutual Migration
Multi-VC Migration Strategies
Live Migration of Homogeneous Virtual Clusters
Live Migration of Heterogeneous Virtual Clusters
Master Node Migration vs. Slave Node Migration
Related Work
Conclusion & Future Work
GRID 2012
Sep. 20-23, 2012 Beijing, China
4. Motivations
Currently, live migration of virtual machine
s has become a key ingredient behind the
management activities of cloud computing
system to achieve the goals of:
Load balancing
Energy saving
Failure recovery
System maintenance
…
GRID 2012
Sep. 20-23, 2012 Beijing, China
5. Motivations
Virtual Cluster is a group of virtual machines conf
igured for a common purpose, such as high perfor
mance computing or parallel computing, with asso
ciated
Storage resource
Operating system
Software environment
Communication protocol
Network configuration
Two notable features of virtual cluster
Large scale
Intensive communication
GRID 2012
Sep. 20-23, 2012 Beijing, China
6. Motivations
Live migration of virtual clusters (VC) faces s
everal new challenges:
Huge Amount of Data
Limitation of Network Bandwidth
Intensive Communication between VMs
Synchronous Latency
Complex VC Migration Strategies
It is necessary to study this new migration sc
enario to investigate the overheads & bottlen
ecks. GRID 2012
Sep. 20-23, 2012 Beijing, China
7. Contributions
Describe a framework VC-Migration to co
ntrol the migration of virtual clusters.
Perform a series of experiments to unders
tand the performance bottleneck and over
heads of virtual cluster migration:
Performance characterization of virtual cluster;
Dynamic migration strategies of virtual cluster
s;
Migration of multiple virtual clusters;
GRID 2012
Sep. 20-23, 2012 Beijing, China
8. Outline
Motivations
VC-Migration Framework
Virtual Cluster Performance
Cross-Domain Virtual Cluster
Virtual Cluster Scalability
Dynamic Migration Strategies of Virtual Cluster
Effects of Memory Configuration
Migration Scalability
Concurrent Migration with Various Granularities
Mutual Migration
Multi-VC Migration Strategies
Live Migration of Homogeneous Virtual Clusters
Live Migration of Heterogeneous Virtual Clusters
Master Node Migration vs. Slave Node Migration
Related Work
Conclusion & Future Work
GRID 2012
Sep. 20-23, 2012 Beijing, China
11. VC-Migration Framework
VC-Migration Strategies
a. Concurrent
Migration with
Various
Granularity
b. Mutual
Migration
(a) (b) c. Homogeneous
Multi-VC
Migration
d. Heterogeneous
Multi-VC
Migration
(c) GRID 2012
(d)
Sep. 20-23, 2012 Beijing, China
12. Outline
Motivations
VC-Migration Framework
Virtual Cluster Performance
Cross-Domain Virtual Cluster
Virtual Cluster Scalability
Dynamic Migration Strategies of Virtual Cluster
Effects of Memory Configuration
Migration Scalability
Concurrent Migration with Various Granularities
Mutual Migration
Multi-VC Migration Strategies
Live Migration of Homogeneous Virtual Clusters
Live Migration of Heterogeneous Virtual Clusters
Master Node Migration vs. Slave Node Migration
Related Work
Conclusion & Future Work
GRID 2012
Sep. 20-23, 2012 Beijing, China
13. Virtual Cluster Performance
Experimental Configuration
Virtual Cluster Configuration
Dell T710 Server, with 2 Quad-core 64bit Xeon processors a
nd 32 GB DRAM.
CentOS 5.6 with kernel version 2.6.18-238.12.1.e15xen in D
omain 0, and Xen 3.3.1 as the hypervisor.
VM (Guest OS) with Ubuntu 8.10, 1 VCPU & 256 MB DRAM.
MPI version is MPICH 2.1.0.8
All the VM images are stored on a separate NFS storage ser
ver
Benchmarks
HPC Challenge Benchmark Suite (HPCC) as the virtual clust
er workloads: HPL, DGEMM, STREAM, PTRANS, RandomA
ccess, FFT, Communication bandwidth and latency
Virt-LM Benchmark: to measure the migration performance a
nd overheads. [Huang et al., ICPE2011]
GRID 2012
Sep. 20-23, 2012 Beijing, China
25. Experimental Findings
The main contradiction of VC migration is the large amount of image
data and the limited network bandwidth.
The virtual machines belonging to the same virtual cluster should be
deployed together as far as possible to reduce the communication a
nd synchronization latency across different physical machines.
Virtual cluster has good scalability and is suitable for the high perfor
mance computing and parallel computing tasks.
When a virtual cluster needs to be migrated, it is important to select
a suitable concurrent migration granularity. Large concurrent granul
arity will decrease the VC performance dramatically.
Mutual migration should be avoided due to the long overall migratio
n time.
The migration of slave node incurs relatively less overhead compar
ed to the master node. So it should give a priority to the slave migrat
ion.
Migration order is important when multiple virtual clusters need to b
e migrated. The long-time cross-domain virtual cluster will decrease
the overall performance of applications.
GRID 2012
There is a big optimization20-23, 2012 in the live migration of virtual clust
Sep. space Beijing, China
26. Outline
Motivations
VC-Migration Framework
Virtual Cluster Performance
Cross-Domain Virtual Cluster
Virtual Cluster Scalability
Dynamic Migration Strategies of Virtual Cluster
Effects of Memory Configuration
Migration Scalability
Concurrent Migration with Various Granularities
Mutual Migration
Multi-VC Migration Strategies
Live Migration of Homogeneous Virtual Clusters
Live Migration of Heterogeneous Virtual Clusters
Master Node Migration vs. Slave Node Migration
Related Work
Conclusion & Future Work
GRID 2012
Sep. 20-23, 2012 Beijing, China
27. Related Work
Virtual Cluster Performance Analysis
MPI Cluster [Huang et al., ICS’06; Ye et al., HPCC’10; Merge
n et al., SIGOPS Oper. Syst. Rev., 2006]
MapReduce Cluster [Ibrahim et al., CloudCom’09]
Live Migration of Single VM
Pre-copy technique [Clark et al., NSDI’05; Nelson et al., US
ENIX’05]
Post-copy technique [Hines et al., VEE’09]
Whole system migration [Luo et al., Cluster’08]
Live Migration of Multiple VM
Resource reservation method [Ye et al., Cloud’10]
Avoid the data de-duplication for concurrent migraiton
[Deshpande et al., HPDC’11; Al-Kiswany et al. HPDC’11]
However, all the above work didn’t solve the problem of live migration of virtual
GRID 2012
cluster in which the intensive Sep. 20-23, 2012 Beijing, China
communication overheads can affect the migration
performance.
28. Outline
Motivations
VC-Migration Framework
Virtual Cluster Performance
Cross-Domain Virtual Cluster
Virtual Cluster Scalability
Dynamic Migration Strategies of Virtual Cluster
Effects of Memory Configuration
Migration Scalability
Concurrent Migration with Various Granularities
Mutual Migration
Multi-VC Migration Strategies
Live Migration of Homogeneous Virtual Clusters
Live Migration of Heterogeneous Virtual Clusters
Master Node Migration vs. Slave Node Migration
Related Work
Conclusion & Future Work
GRID 2012
Sep. 20-23, 2012 Beijing, China
29. Conclusion
We have studied the live migration perfor
mance and overheads of virtual clusters fr
om the experimental perspective and inve
stigated different VC migration strategies.
Experimental results reveal some new dis
coveries, based on which we can propose
several optimization principles to improve t
he migration performance of virtual cluster
s.
GRID 2012
Sep. 20-23, 2012 Beijing, China
30. Future Work
Optimize the migration mechanism in the
hypervisor to improve the migration efficie
ncy of virtual clusters
Design efficient migration algorithms for th
e virtual clusters.
GRID 2012
Sep. 20-23, 2012 Beijing, China