Skip to content

Use of a Nvidia K40 in OLB

Due to recent bot attacks we have changed the sign-up process. If you want to participate in our forum, first register on this website and then send a message via our contact form.

› Forums › Lattice Boltzmann Methods › General Topics › Use of a Nvidia K40 in OLB

Viewing 9 posts - 1 through 9 (of 9 total)
  • Author
    Posts
  • September 5, 2018 at 11:38 am #1992
    LaurentDelaon
    Participant

    Hi Mathias,

    How can I verify the usage of the K40 by OpenLB?

    I have a doubt about the use of the K40 GPU with openLB.
    I have put the parallel MPI (and the MPIC++) option on in the base directory.
    But when I use the nvidia-smi it’s seem not use it…
    I use a UbuntMate14.04Lts.
    How can I do it ?

    Regard’s

    Laurent.

    September 5, 2018 at 2:58 pm #2923
    Markus Mohrhard
    Participant

    Hello Laurent,

    the currently released OpenLB versions don’t support GPUs yet. We have an unreleased prototype that is using CUDA for GPU support and hope to release initial GPU support as part of OpenLB 1.3.

    Regards,
    Markus

    September 5, 2018 at 4:02 pm #2924
    LaurentDelaon
    Participant

    HI Markus,
    ok I waiting for it!
    I have imagined that mpi can benefit of th K40…Seem not true.
    A question : with mpi I benefit better way to use my processors.
    I have 31cores but this is work better with -np 15(wich give the simulation quicker).
    How explain that ?
    machine use a XEON E5-2620v4@2.1Ghz with 31 cores

    September 5, 2018 at 4:22 pm #2925
    Markus Mohrhard
    Participant

    Hey Laurent,

    most likely you have 2 E5-2620v4 which are 8 core/16 thread CPUs. In total you will see 32 virtual cores but there are only 16 physical cores, so the best performance will be reached with 16 MPI jobs. With anything above 16 you have additional context switches that take time and limit the available cache per core.

    For additional information about that I recommend to read about Simultaneous multithreading (SMT) or Hyper-threading (HT).

    Regards,
    Markus

    September 6, 2018 at 10:48 am #2926
    LaurentDelaon
    Participant

    Thank’s Marcus.

    September 17, 2018 at 1:59 pm #2927
    LaurentDelaon
    Participant

    Hi Marcus,

    is it possible to test pre-release 1.3 with GPU for a K40 ?
    I only use cylinder2d base example.

    Laurent.

    September 17, 2018 at 5:07 pm #2928
    Markus Mohrhard
    Participant

    Hello Laurent,

    yes it will be possible as soon as it is in a reasonable shape. Please contact Mathias Krause to find out how and when we can provide you with a pre-release tarball. Note that the work is currently still ongoing and will hopefully be merged into our main development branch in the next weeks.

    Regards,
    Markus

    June 29, 2021 at 3:15 pm #5760
    POSThu
    Participant

    Hi Markus,

    I wonder whether the latest OpenLB supports the GPU computation now? I want to further accelerate the simulation process to achieve real-time simulation of airflow field.

    Best Regards,
    mengqiang

    July 1, 2021 at 10:45 am #5768
    Adrian
    Keymaster

    No, unfortunately the merge in 2018 did not work out due to unforeseen problems. However I am currently working on large changes in OpenLB to support both vectorization on CPUs and utilization of GPUs that will be published in release 1.5.

  • Author
    Posts
Viewing 9 posts - 1 through 9 (of 9 total)
  • You must be logged in to reply to this topic.