Distributed Computing in Practice: The Condor Experience

D. Thain, T. Tannenbaum, and M. Livny.
University of Wisconsin

One-line Summary

Overview/Main Points

• Architecture outline
  • User writes ProblemSolver to describe jobs and their requirements, which is submitted to Agent.
  • Resource provides computation environments for executing jobs within Sandbox.
  • Matchmaker negotiates between Agent and Resource for matched job requirements.
  • Agent forks Shadow to communicate with Sandbox in terms of job execution.
• Architecture evolution
  • A Condor pool
  • Gateway flocking
  • Direct flocking
  • Condor-G(rid)
    • initial design
    • personal Condor pool
• Condor in details
  • Planning & Scheduling
    • planning
    • scheduling
    • ClassAds: Job and Machine ClassAds
    • planning within a schedulre: make choices about where to submit and when to submit.
    • scheduling within a plan
      • suppose where and when are fixed
      • create a schedule for all the tasks
  • ProblemSolver: a high-level parallel design pattern
    • Master-worker
      • Workder process runs jobs
      • Master process steers among workList and tracks worker processes
    • DAG
  • Split execution: conduct secure remote execution
    • shadow
    • sand-box
    • the standard universe
      • two-phase open for handling I/O requests
      • directing system calls to the original Agent node
  • Security issues
    • Secure communication: secure RPC
    • Secure execution: user-id based mgmt
  • Reliablity issues: error handling
    • what if a remote server suddenly shuts down? Contact MatchMaker for another Resource.
    • what if the Condor job itself produces errors? Automatically re-schedule

Relevance

Flaws