Slide 1: Debian GNU/Linux 3.1 (Sarge) at GSI

Christopher Huhn

GSI IT department, C.Huhn@gsi.de

Slide 2: Woody installation

• Woody installation 2002/2003
• Server installation: cloning over NFS
  • One size fits all

Slide 3: Woody installation

• Configuration management: shell scripts
  • Who dunnit? No revision control, no one to blame
• Centralisation achieved by central NFS mount /usr/local
• Sparse documentation
  • Know-how often only implicit
• Administration techniques don't scale well
  • particularly in relation to the number of admins

Slide 4: Woody client

• "Groupserver" concept:
  • classic NFS root concept
  • shared read-only /usr
  • Client system - almost - completely on server (except for /var and /tmp)
• Old server hardware
  • Hardware maintenance painful
• Important binaries writeable (i.e. /sbin/init)
• No automatic upgrade for anything outside /usr
  • I. e. no automatic security updates for /bin, /sbin binaries
  • No automatic distribution of /var or /etc files

Slide 5: Into the future ...

• http://www.infrastructures.org/ - Site management best practices
  • System administration should not work like craftsmen
  • Infrastructure architects
  • Configuration as condensed and central as possible
• FAI: Fully Automatic Installation framework
  • Plan your installation and FAI installs your plan
  • Debian-centric but very flexible

Slide 6: Into the future ...

• Cfengine - configuration management framework
  • centralised and pull-based
  • description of the intended configuration state
  • not a sequential list of actions to be performed
• Use of Debian tools (Package management, Debconf) wherever appropriate
• Debian repositories used:
  • Debian, Debian-Security, Backports.org, Debian volatile, GSI repository

Slide 7: Into the future ...

• More services - increased diversification
• Divide installation and configuration into host classes
• Keep It Small and Simple - only install what's required for a service
• Installation and configuration system excellence indicator:
  • Reinstallation to service production state should be
    • completely unattended and
    • faster than a restore from tape backup

Slide 8: ... back to the past?

• NFS root reinvented:
  • Single system image
    • Shared by all clients
    • managed with cfengine
    • completely distinct from the server system
  • read-only root filesystem
  • unionfs filesystem
  • All host-specific files in a central Subversion repository
    • Accessed via WebDAV/davfs

Slide 9: Pros

• Improved security
  • Difficult (impossible?) to hack
  • Updates even when the box is down
    • ready for dual boot / multi boot
• Improved reliability?
• Improved performance (new hardware)
• Instant installation (~ 2 minutes)
  • Can be performed by the operators
• Ease of administration
  • Condensed and centralised

Slide 10: ... and Cons

• Completely reliant on the network
• Inferior performance compared to local installation?
• Installation time irrelevant if installation request response time is in size of days?

Slide 11: On and on and on and on ...

Slide 12: Fabric management

• Set ALLOWTOPICCHANGE = Christo
• Set ALLOWTOPICVIEW = *