locked
cuda and remote dev RRS feed

  • Question

  • Hi, we are considering running Windows HPC 2008 on a compute node with a Tesla S1070.

    So far, all of our CUDA development has been on XP 32-bit with single GPU (GTX 285 or Quadro 4800).

    We have noticed that the CUDA apps in general won't run over XP's Remote Desktop.  I found some websites which describe the problem and say that options such as VNC must be used as an alternative.

    I am wondering if Windows HPC will solve my problem.  I'd like to be able to run my GUI and command-line CUDA apps over remote desktop.

    Thanks in advance.
    Friday, December 18, 2009 10:28 PM

Answers

  • Hi,

    The reason CUDA apps will generally not work correctly via remote desktop is that, in general, the graphics device driver makes the GPGPU device accessible only when the application is running on the interactive console (i.e. when the user is logged into the console session connected to the physical keyboard/video/mouse). When an application is instead running on a console logged in via remote desktop, the graphics device driver generally does not have access to the physical GPGPU device.

    If you would like to use the current version of Windows HPC to run CUDA applications, then you can do so by making sure that the correct user is logged into the interactive console on the GPGPU compute node. When that user submits their Windows HPC job containing CUDA tasks, if they set the job environment variable HPC_ATTACHTOCONSOLE=TRUE, then any CUDA task in the job will run on the logged-in interactive console and therefore have access to the GPGPU. We are also actively working on additional solutions for the next Windows HPC version (V3) to make it easier to run CUDA code on a Windows HPC cluster with GPGPUs.

    Regards,

    Patrick
    Friday, December 18, 2009 11:26 PM