A Research Project Idea

• Looking through SC 2018, I found this paper: Characterization of MPI Usage on a Production Supercomputer
  • MPI: Message Passing Interface
    • Still the library for distributed computing.
  • This is a fairly easy library to learn.
  • The paper does the following
    • Introduces the problem.
      • Not so much of a lit review, just a statement of the problem.
      • Lists the results
      • Says why the results are important.
    • Reviews a tool they created for analyzing parallel computations.
    • Detailed analysis of the jobs that were monitored
      • Includes details on how jobs were eliminated from consideration.
      • This, along with the next section is the main point of the article.
      • A presentation of the data and methods.
      • This is really where you judge the paper.
    • Analysis of the MPI calls used
    • Related work
    • Conclusions.
  • This is missing a future work section.
• But, Could I do something with this as a project?
  • Is the autoperf tool available?
    • Yes it is.
  • Is MPI available?
    • Yes it is.
  • Are there any example MPI programs I could test?
    • Yes there are
• Outline of research
  • Question:
    • Can I reproduce the conditions of the paper? - no but valid work
    • Do example programs behave like the larger scientific programs?
    • Better: Is there a major difference between Grand Challenge computations and example MPI programs?
  • Methods (or work to do)
    • Read and understand the paper (so you know what you are trying to do)
    • Get MPI and autoperf set up on a small cluster of machines.
    • Work through a MPI tutorial.
    • Get example programs running.
    • Perform analysis of example programs.
  • Possible problems
    • We don't have an MPI cluster, so we would have to use lab machines.
    • We could use euler as well, but results would be highly skewed.