No announcement yet.

InfiniBand support

  • Filter
  • Time
  • Show
Clear All
new posts

  • InfiniBand support

    is InfiniBand support planned for the future?
    This would be very helpful to stress InfiniBand adapters and the corresponding switches.

    This seems to be a good resource currently, they also offer various IB benchmarks with sources:

    Side question:
    Do you know if the DirectCompute benchmark runs with Tesla cards? These don't have a video output.
    (Same for BurnInTest)

  • #2
    While there is no direct InfiniBand test, there are tests that can make indirect use of InfiniBand.

    For example if you are running Ethernet over InfiniBand, then you can use the networking test.
    If you are SCSI disks over InfiniBand, then you can use the disk test.


    • #3
      While strictly not necessary, the DirectCompute test include functionality to visualize the results of the Direct Compute calculations.

      The visualizations expect to have some video output available. So I would guess these would fail to start up if there was no video card available. But I haven't tested the exact behaviour.


      • #4
        Thanks David.
        I am not using Ethernet over InfiniBand, so it doesn't seem very useful for this test area as the configuration wouldn't match the productive environment.
        Will have a closer look at the mvapich when I got some free time, maybe it can be applied in an easy way.

        There is a primary video output on the server, but I am not sure how the benchmark uses them?
        Example: We use the internal GPU and have two Tesla cards installed for GPU computing.
        Can the test cover that? (Actually I am not even interested in the performance of the Intel GPU, as long as it provides an image I am fine with it, it's for administration purposes only)


        • #5
          It should be picking the 'primary' video card for the test and the display of the calculation results. But primary has become a vague concept in the last few year. SLI, Crossfire, Optimus, and more unusual setups like yours all confuse the situation.


          • #6
            It's actually the normal way to use Tesla cards.
            I don't think there exists one of those with a video output at all.
            They are designed for computing purposes only.

            To make this working the test would have to explicitly use the additional Tesla cards, as none of them is the primary device and redirect the output (which could be a bottleneck).
            You have to threat Tesla cards like co-processors (similar to Intel Xeon Phi) as far as I am aware.
            I assume the tests don't run with Xeon Phi cards either or is it worth a shot there?


            • #7
              It's actually the normal way to use Tesla cards.
              Probably is.
              But my point is that is isn't a typical PC setup. Maybe 0.001% of machines are setup like yours.

              My understanding is that these cards are near useless for running normal software (games, Photoshop, video editing, etc..). But they come into their own when custom code is written to exploit them. And we haven't written this type of custom code.

              Same deal for Phi cards.

              Of course it is technically possible to write code for them, but it is an expensive exercise to buy all the exotic hardware for development and testing, when the market share for this hardware is near zero. We would be happy to do it as a paid consulting job however.