???????????????????????????????????????? Xen, ROCKS, ?????????? Roll Implementation of Virtualization Clusters based on Xen, ROCKS, and ThaiGrid Roll

1
???????????????????????????????????????? Xen,
ROCKS, ?????????? RollImplementation of
Virtualization Clusters based on Xen, ROCKS, and
ThaiGrid Roll

• By
• Supakit Prueksaaroon1, Wittaya Konghaped2, Vara
  Varavithya2, and Sornthep Vannarat1
• 1Large-scale Simulation Research Laboratory,
• National Electronics and Computer Technology
  Center
• 2Department of Electrical Engineering,
• Faculty of Engineering, King Mongkuts Institute
  of Technology North Bangkok
• Present by
• Wittaya Konghaped
• Department of Electrical Engineering,
• Faculty of Engineering, King Mongkuts Institute
  of Technology North Bangkok

2
Agenda

• Introduction to Virtualization technology
• ROCKS Cluster ThaiGrid Roll
• Virtual Cluster
• Performance Measurement of Virtual Cluster
• Conclusion

3
Virtualization Concepts

• A further virtualization layer in the middle
  between OS and HW Virtual Machine Monitor (VMM)
• Allow for multiple concurrent OS instances
• Modern PCs are powerful enough for creating the
  illusion of several OS virtual machines to run
  simultaneously

Apps
Apps
Apps
Apps
OS
OS
OS
OS
VMM
HW
HW
4
Introduction to VM technology

• Virtual Machine have 3 technologies
• Emulation, full system simulation, or "full
  virtualization with dynamic recompilation" the
  virtual machine simulates the complete hardware,
  allowing an unmodified OS for a completely
  different CPU to be run.
• Native virtualization and "full virtualization"
  the virtual machine only partially simulates
  enough hardware to allow an unmodified OS to be
  run in isolation, but the guest OS must be
  designed for the same type of CPU. The term
  native virtualization is also sometimes used to
  designate that hardware assistance through
  Virtualization Technology is used.
• Paravirtualization the virtual machine does not
  simulate hardware but instead offers a special
  API that requires OS modifications.

From http//en.wikipedia.org/wiki/Virtual_machine
5
Full virtualization

• There is a complete functional ordering between
  layers
• Full abstraction of machine (from BIOS to disks,
  DMA controlers, video)
• Virtualization is fully transparent guest OS
  unchanged
• Much more complex to design and implement

6
Paravirtualization

• Not a really hierachical ordering between layers
• Virtualization is similar to HW interface but
  neither complete nor identical
• Guest OSs must be modified to become VM-aware
• There is a potential gain in performance, due to
  specialization of kernel code
• Easier to design
• But carefully think about interfaces

7
Benefit of Virtualization for HPC

• Better utilize Hardware resources
• Independent to custom library
• Easy to create/destroy the guest OSs
• The key capabilities for
• High Availability computing resources
• Migration
• The perfect sandbox

8
ROCKS

• The tools for create the Cluster computer.
• Front-node
• Compute-nodes
• ROCKS ROLL architecture
• Features
• Cluster tools for management cluster system
• Support multiple software for HPC and Grid
• Bio-roll, Grid-roll, Intel-roll, Viz-roll and so
  on
• Easy re-build and install compute nodes

9
ThaiGrid Roll

• The complete package of ROCKS and integrated Grid
  packages for support Grid environment in
  Thailand.
• ThaiGrid Monitoring tools
• ThaiGrid CA
• ThaiGrid scripts
• Developed by Thai National Grid Project

10
Problems of Native Cluster

• Application porting
• Library incompatibility
• OS Software version incompatibility
• Security
• Configuration complexity
• High administrative cost
• Heterogenious operating system support

11
In this work

• Initial Study to implement Virtual Cluster
• We implement VM based on ThaiGrid Roll
• Objectives
• To compare overall performance and get the
  measurement for large scale Virtual cluster
  simulation.
• For ThaiGrid Setup can make use of CPU times for
  others task such as information Server, DB Server
  and Co-exist well with Grid Services

12
Goal

• Shared Grid resources and running concurrent with
  production resources (Web, Mail, DB server, etc.)

13
Definition

• Definition of this work
• Domain-0 or (Dom-0) is the based OS that Virtual
  machines are running on
• Domain-U or (Dom-U) is the guest OS that are
  running simultaneously with domain-0

Domain-U
Apps
Apps
Apps
OS
OS
OS
VMM
Domain-0
HW
14
Implement concept

• Build the ThaiGrid Roll and ROCKS 4.2.1
• Build the prototype Front-end node Compute node
• Install Dom-U image (kernel-xenU
  2.6.16-xen3_86.1_rhel4.1)
• Copy Data to image file
• Setup the Dom-0 machine by using Fedora Core 5
• Distributing the images to front-end and all
  compute nodes

Compute
Front-end
Compute
15
Images Maker

• Create the copy of prototype compute node image
  configuration files
• mount o loop ltIMAGEgt ltMount-ponitgt
• Edit /etc/hosts, /etc/sysconfig/network,
  /etc/sysconfig/network-script/ifcfg-eth0 and so
  on
• umount ltMount-pointgt
• Update SGE or PBS
• xm create c config

16
Experiment Details

• 3 nodes of Satellite Cluster
• IBM x336 Dual Xeon processors 2.8GHz
• RAM 4GB
• Harddisk SCSI 73GB
• Intel e1000 network interface
• Inter-connection 10/100 Mbps
• HPL benchmark

17
HPL Benchmarks
18
Parallel Becnhmark
19
Conclusion

• The result show the performance nearly the native
  linux in case of internal machine
• High gap-performance in case of parallel machines
  
• Virtual Cluster should be based on
  High-Throughput Computing
• Xen should be improved the I/O performance
• This result of this work use for initial
  implement of Virtual Cluster.
• Virtualization technology show high potential for
  HPC

20
Future Work

• Simulate large-scale virtualization based on this
  works and projection performance
• Scheduling move to Sandbox scheduling than job
  scheduling
• Security issues
• Compatibility issues
• Investigate high throughput application
  performance on virtualization technology
• Formulate virtualization efficiency for
  computation intensive scheduling

21
Acknowledgement

• Thank you ThaiGrid for support the cluster
  machines.

22
Thank you Question
23
HPL CPU context switch overhead
HPL parameter N9000 NB64
24
Percentage of CPU context switch overhead
HPL Overhead 0.66(NumberDomain-U) SPECCPU2000
Overhead 0.77(NumberDomain-U)