Autonomic Live Adaptation of Virtual Computational Environments in a Multi-Domain Infrastructure Paul Ruth, Junghwan Rhee, Dongyan Xu Department of Computer Science and Center for Education and Research in Information Assurance and Security

1
Autonomic Live Adaptation of Virtual
Computational Environments in a Multi-Domain
InfrastructurePaul Ruth, Junghwan Rhee,
Dongyan XuDepartment of Computer Science and
Center for Education and Research in Information
Assurance and Security (CERIAS)Rick Kennell,
Sebastien GoasguenRosen Center for Advanced
ComputingPurdue UniversityWest Lafayette,
Indiana, USA
IEEE International Conference on Autonomic
Computing (ICAC06)
2
Outline of Talk

• Motivations
• Overall architecture
• Design and implementation
• Real-world deployment in nanoHUB
• Related work
• Conclusion
• Demo

3
Motivations

• Formation of shared distributed
  cyberinfrastructure (CI)
• Spanning multiple domains
• Serving users/user communities with diverse
  computation needs
• Exhibiting dynamic resource availability and
  workload
• Need for virtual distributed environments
  (VIOLINs), each with
• Customizability and legacy application
  compatibility
• Administrative privileges
• Isolation, security, and accountability
• Autonomic adaptation capability
• - A unique opportunity brought by
  virtualization (VMs and VNs)

4
Adaptive VIOLINs
Virtual clusters (VIOLINs)
Physical cluster
nanoHUB infrastructure_at_Purdue
5
Autonomic VIOLIN Adaptation

• Adaptation triggers
• Dynamic availability of infrastructural
  resources
• Dynamic resource needs of applications running
  inside
• Adaptation actions
• Resource re-allocation
• Scale adjustment (adding/deleting virtual
  machines)
• Re-location (migrating virtual machines)
• Adaptation goals
• Improving application performance
• Increasing infrastructural resource utilization
• Maintaining user/application transparency
• Minimizing infrastructure administrator attention

6
Research Challenges

• Autonomic live adaptation mechanisms
• VM Resource monitoring and scaling
• Application profiling and non-intrusive sensing
  of application needs
• Live VIOLIN re-location across domains
• Adaptation policies
• VIOLIN adaptation model
• Infrastructure resource availability and
  topology
• Application resource needs
• Application configuration and topology
• Optimal VIOLIN adaptation decision-making
• Goals (cost vs. gains)?
• When to adapt?
• How and how much to adapt?
• Where to migrate?

7
Overall Architecture
VMs
VMs
VIOLIN Switch
VIOLIN Switch
VIOLIN Switch
VIOLIN Switch
Monitoring Daemon
Monitoring Daemon
Dom0
Dom0
VMM
VMM
VMs
VMs
Physical Network
VIOLIN Switch
VIOLIN Switch
VIOLIN Switch
CPU Update
Adaptation Manager
Monitoring Daemon
Monitoring Daemon
Dom0
Dom0
Scale Up
Migrate
VMM
VMM
8
VIOLIN Adaptation Policies

• Maintain desirable resource utilization level
• Reclaim resource if under-utilized over a period
• Add resource if over-utilized over a period
• Scale up local resource share
• Migrate to other host(s)
• Balance host workload
• Intra-domain migration first
• Minimize migration
• Re-adjust resource according to application
  needs

9
Implementation and Deployment

• Extension to non-adaptive VIOLIN
• Based on Xen 3.0 (w/ VM Live migration
  capability)
• Enabling live VIOLIN migration across domains
• IP addresses of VMs
• Root file systems of VMs
• Leveraging Xen libraries for VM resource
  monitoring (xenstat, xentop)
• Extending VIOLIN switch for inter-VM bandwidth
  monitoring
• Deployment in nanoHUB
• On-line, on-demand simulation service for
  nanotechnology community
• Web interface for regular users
• My workspace interface for advanced users
• Local infrastructure two clusters in two
  subnets

10
nanoHUB Deployment Overview
Local Virtual Machines Migratable Isolated from
Local infrastructure
VIOLIN Virtual Cluster
Delegated trust
Virtual Infrastructure over WAN
11
VIOLIN in nanoHUB
In the backround
VIOLIN
Simulation job
12
VIOLIN in nanoHUB
13
Impact of Migration on App. Execution
End-to-end execution time of NEMO3D w/ and w/o
live VIOLIN migration
14
VIOLIN Adaptation Scenario
1. Initially VIOLIN 1, 2, 3 are computing, VIOLIN
2 is about to be finished.
2. After VIOLIN 2 is finished, before adaptation
Without Adaptation
With Adaptation
15
VIOLIN Adaptation Scenario
2. After VIOLIN 2 is finished, before adaptation
3. After adaptation
Without Adaptation
With Adaptation
16
VIOLIN Adaptation Scenario
4. After VIOLIN 4, 5 are created
3. After adaptation
Without Adaptation
With Adaptation
17
VIOLIN Adaptation Scenario
4. After VIOLIN 4, 5 are created
5. After VIOLIN 1, 3 are finished
Without Adaptation
With Adaptation
18
VIOLIN Adaptation Scenario
6. ALL VIOLINs are finished
5. After VIOLIN 1, 3 are finished
Without Adaptation
With Adaptation
19
Limitations and Future Work

• Simple, heuristic adaptation policy
• Application of machine learning and data mining
  techniques
• Centralized adaptation manager
• Hierarchical or peer-to-peer adaptation managers
• Imprecise application resource demand inference
• Multi-dimensional, fine-grain resource demand
  profiling
• Campus-wide infrastructure
• Evaluation and deployment in wide-area
  infrastructure

20
Related Work

• VNET (Northwestern U.)
• Cluster-on-Demand (COD) (Duke U.)
• Virtual Workspaces on Grid (Argonne National
  Lab)
• SoftUDC (HP Labs)
• WOW and IPOP (U. Florida)

21
Conclusions

• Autonomically adaptive virtual infrastructures
  (VIOLINs)
• A new opportunity brought by virtualization
  technologies
• Decoupled from underlying shared infrastructure
• Intelligent, first-class entities with
  user-transparent resource provisioning
• Key benefits
• Application performance improvement
• Infrastructure resource utilization
• Management convenience (at both virtual and
  physical levels)

The Cray motto is adapt the system to the
application - not the application to the
system. - Steve Scott, CTO, Cray
Inc. on adaptive supercomputing, March 2006
22
Thank you.
For more information Email dxu_at_cs.purdue.edu U
RL http//www.cs.purdue.edu/dxu Google
Purdue VIOLIN FRIENDS