Transcript cooperative linux

Reading
CampusGrid
e-Science Institute - Edinburgh
High Throughput Computing Week
29 November 2007
© University of Reading 2007
www.reading.ac.uk
Presenter
Christopher Chapman
Systems Programmer (ITNG)
School of Systems Engineering
University of Reading
Overview
• Designing the CampusGrid
• Current Implementation
• Success Stories
• The Future – CampusGrid 2.0
Background
• Started as a School of Systems Engineering activity – then
needed to become campus wide
• Condor Implementation
• Different Implementations
– Diskless Condor – Worked but was restricted by memory
– Dual booting Windows/Linux
• Problems
–
–
–
–
Limited nodes
Only available overnight
Rebooting interfered with scheduled maintenance
Not scalable
Designing the CampusGrid
• Requirements
–
–
–
–
–
Needs to be scalable
Needs to be Linux based
Require minimal changes to infrastructure
Must not impact normal PC operation/maintenance
Needs to be deployable though existing mechanisms
• Solution
–
–
–
–
Virtual instance of Linux running under Windows
Network configuration based on host
Only active when no one is logged into Windows
Packaged for Microsoft SMS deployment
Virtualisation
• VMware
• Virtual PC
• Cooperative Linux (coLinux)
Virtualisation - VMware
• Potential benefits
–
–
–
–
–
Well established
Support for Linux including performance additions
Easily modifiable configuration file
Easy to deploy
Results in fully native Linux install
• Potential challenges
– Requires server version to run as a service
– Hard to pass information between host and guest
– Not free (at the time!)
Virtualisation – Virtual PC
• Potential benefits
–
–
–
–
Free - Included in Campus Agreement
Easily modifiable configuration file
Straightforward to deploy
Results in fully native Linux install
• Potential challenges
– Requires server version to run as a service
– Hard to pass information between host and guest
– No official Linux support (At the time!)
Virtualisation – coLinux
• Potential benefits
–
–
–
–
–
–
Free
Easily modifiable configuration file
Straightforward to deploy
Easy to pass information between host and guest
Runs as a service
Designed for Linux
• Potential challenges
– Requires custom built Linux install
– Requires 3rd party network drivers
– Only virtualises disk and network – no graphics
And the winner?
COOPERATIVE LINUX
Cooperative Linux
• coLinux kernel runs in a privileged mode
• PC switches between host OS state and coLinux kernel
state
• coLinux as full control of the physical memory
management unit (MMU)
• Almost the same performance and functionality that can
be expected from regular Linux on the same hardware
• Can run a number of different Linux flavours
Network configuration
• Hosts use a Public class B network
• Guests mirror the Hosts IP but for a private class B
Network
e.g. HOST =
134.225.217.32
GUEST =
10.216. 217.32
–
–
–
–
Guest MAC address based on Hosts MAC address
No DHCP
Private IP address are registered with DNS
Each subnet has a second gateway (multi-netting)
Nonintrusive
1. coLinux starts when a user logs out
2. Condor starts excepting jobs
3. coLinux start shuts down when user logs in
4. Condor is shutdown, nicely evicting any jobs
Current Implementation
• Campus wide
• coLinux running Fedora Core 3
• Private network
• Condor
• NGS certificate authentication or Kerberos
• Shared File system - NFS
• Deployed using SMS
• Around 400 nodes
Successes to date
• TRACK - Feature detection, Tracking and Statistical analysis for
Meteorological and
Oceanographic data.
– Processing around 2000 jobs a day
• Computing Mutual Information - Face Verification Using Gabor
Wavelets
– Condor cut processing from 105 days to 20 hours
• Bayes Phylogenies - inferring phylogenetic trees using Bayesian
Markov Chain Monte Carlo (MCMC) or Metropolis-coupled Markov
chain Monte Carlo (MCMCMC) methods.
– Implementation of MPI on condor
• Analysing Chess moves
The Future – CampusGrid 2.0
• coLinux running Linux (version to be decided)
• Private network
• NGS affiliate status
• VDT software stack
• NGS certificate authentication or Kerberos
– Maybe even ShibGrid
• Shared File system - NFS
• Nodes updated from a central server
• Deployed using SMS
• More nodes
For further information
E-mail
Christopher Chapman
[email protected]
Jon Blower
[email protected]
Websites
http://www.resc.rdg.ac.uk/
http://wiki.sse.rdg.ac.uk/wiki/ITNG:Campus_Grid