check that mainstream machines use correct thread binding

Create issue
Issue #1528 new
Roland Haas created an issue

it is possible that systems that initialize themselves before hwloc (eg MPI) bind to the wrong CPU core unless (the equivalent of) numactl is used.

For an example on bluewaters see #1527.

Keyword:

Comments (4)

  1. Erik Schnetter
    • removed comment

    This had a significant performance impact on Blue Waters (from 14 M/hr to 17 M/hr).

  2. Erik Schnetter
    • removed comment

    We need to check the following systems:

    • bluewaters (done)
    • carver
    • datura
    • gordon
    • hopper (done)
    • lonestar
    • mike (done)
    • orca
    • pandora
    • philip
    • queenbee
    • shelob (done)
    • stampede (done)
    • stampede-mic
    • titan (done)
    • trestles
    • vesta
    • zwicky

    Please edit this text, appending e.g. "(done)" when a machine has been checked.

  3. Ian Hinder
    • removed comment

    Can hwloc be given a parameter to abort with an error message if the thread binding is not correct? If it doesn't know what should be correct, this could be passed in using some environment variables. This would simplify the testing.

  4. Erik Schnetter
    • removed comment

    Yes, such a parameter would be possible. It is in general not clear (without experimentation) which thread bindings are optimal, but we can flag the really bad cases where processes are not bound to the correct socket.

  5. Log in to comment