CDF Interactive Login Pool

1
CDF Interactive Login Pool

• What is the Interactive Login Pool?
• A misnomer. This project is intended to develop
  a replacement of CDF's large SMP machine,
  fcdfsgi2, using less costly hardware running
  Linux. Logins are only a subset of the functions
  which should be implemented.
• Why is this desirable?
• Elimination of expensive service contracts for
  SGI/IRIX.
• Binary compatibility with CAF.

2
CDF Interactive Login Pool

• What needs to be done?
• Fcdfsgi2 performs several functions
• Logins (including account creation and
  management)
• Physics group disk shares (including the 'Look
  Area' and access via rootd)
• Batch system

3
CDF Interactive Login Pool

• How can these functions be implemented?
• Batch System
• Batch system on fcdfsgi2 intended for user
  analyses. Since migrated to CAF.
• Physics Group Disk Shares
• 50TB of Fibre Channel disk connected to
  fcdfsgi2. Ideally, migrate disk to the Linux
  cluster. Exploring implementation of SAN using
  Fibre Channel switches connecting disk to nodes
  serving volumes via NFS.

4
CDF Interactive Login Pool

• How? (cont.)
• Logins
• Several possibilities
• Master/Slave Load Sharing (e.g. OpenMOSIX)
• Batch system (e.g. FBSNG on CMS User Analysis
  Farm)
• IP Load Balancing
• Round-robin DNS
• Switch-level load-balancing
• CERN variant of IBM's eNetworkDispatcher
• Linux Virtual Server (LVS)
• Many stand-alone nodes (e.g. Fcdflnx2, 3)
• Fallback, if out of time, or options above fail

5
CDF Interactive Login Pool

• Logins (cont.)
• Trials and Tribulations
• CDF/MIT tested OpenMOSIX cluster
• Highly unstable, prone to downtime due to single
  node failure
• Round-robin DNS
• DataComm unwilling to implement due to
  troubleshooting difficulties
• No means of dealing with node failure
• LVS
• Forwarded TGT is addressed to the shared name
  rather than the server being used. Cannot
  connect to machines FROM LVS. Current work is
  trying to overcome this.

6
CDF Interactive Login Pool

• Logins (cont.)
• Untested options
• Switch-level load-balancing
• Need more information from network experts
  regarding support and implementation
• FBSNG
• eNetworkDispatcher variant
• Awaiting information from CERN

7
CDF Interactive Login Pool

• Current directions
• Beginning joint effort with CMS to develop
  mutually beneficial solution
• Searching for KerberosLVS workaround
• Inquiring into CERN solution
• Testing unused storage hardware with Fibre
  Channel switch
• Account management via NIS cluster
• Bottom line
• Solutions in place by end of 2004