CUDA on Linux


Advanced search

Message boards : Cruncher's Corner : CUDA on Linux

AuthorMessage
Profile microchip
Send message
Joined: Jun 10 06
Posts: 48
Credit: 6,120,995
RAC: 4,035
Message 111676 - Posted 15 Apr 2011 16:57:25 UTC

    Hi,

    Got a question here. I'm fairly new to CUDA and would like to add a CUDA capable GPU in my crunch system. I'm on Linux and my question is if I have to install anything special to make it work with BOINC and E@H? I mean do I need the proprietary driver from nVIDIA and do I need some specific CUDA kit for it to work?

    Thanks :)
    ____________

    Team Belgium

    Michael Karlinsky
    Avatar
    Send message
    Joined: Jan 22 05
    Posts: 896
    Credit: 14,437,629
    RAC: 20,594
    Message 111677 - Posted 15 Apr 2011 17:18:30 UTC - in response to Message 111676.

      Hi,

      you need to install the NVIDIA driver.

      download

      and

      instructions

      excerpt:


      1) Kernel sources must be installed and configured. Usually this means
      installing the 'kernel-source', 'make' and 'gcc' packages with YaST2.

      2) Run the following commands

      echo "blacklist nouveau" > /etc/modprobe.d/nvidia.conf
      # recreate initrd without KMS, if the use of KMS is enabled in initrd
      if grep -q NO_KMS_IN_INITRD=\"no\" /etc/sysconfig/kernel; then
      sed -i 's/NO_KMS_IN_INITRD.*/NO_KMS_IN_INITRD="yes"/g' /etc/sysconfig/kernel
      mkinitrd
      fi

      3) Reboot your machine.

      4) Use the nvidia installer for 260.19.36.

      sh NVIDIA-Linux-x86_64-260.19.36.run -q

      IMPORTANT: You need to recompile and install the nvidia kernel module
      after each kernel update.

      sh NVIDIA-Linux-x86_64-260.19.36.run -K



      Michael

      PS: instructions are for SUSE, but I guess they should work everywhere else.
      ____________
      Team Linux Users Everywhere

      Profile microchip
      Send message
      Joined: Jun 10 06
      Posts: 48
      Credit: 6,120,995
      RAC: 4,035
      Message 111678 - Posted 15 Apr 2011 17:39:57 UTC

        Thanks for the info. Yes, I know how to manually install the NV driver, but thanks anyways... So I don't need anything else for CUDA on linux? I saw on the NV site a CUDA kit but it's an SDK, not sure if I need that
        ____________

        Team Belgium

        Michael Karlinsky
        Avatar
        Send message
        Joined: Jan 22 05
        Posts: 896
        Credit: 14,437,629
        RAC: 20,594
        Message 111679 - Posted 15 Apr 2011 18:01:16 UTC - in response to Message 111678.

          I saw on the NV site a CUDA kit but it's an SDK, not sure if I need that


          This is for development only. All BOINC-project (should) provide the rest (there are some libs, which are necessary).

          Michael

          ____________
          Team Linux Users Everywhere

          Jeroen
          Send message
          Joined: Nov 25 05
          Posts: 326
          Credit: 384,988,515
          RAC: 569,399
          Message 111687 - Posted 15 Apr 2011 22:51:01 UTC - in response to Message 111676.

            Last modified: 15 Apr 2011 22:51:27 UTC

            Depending on which GPU you add to your system, you may be able to run multiple work units at once per GPU using an app_info.xml file. Specifically the 4xx and 5xx Fermi GPUs can run two or more work units at a time optimally provided there is enough GPU memory. Each work units requires around 300-350 MB of GPU memory.

            Profile microchip
            Send message
            Joined: Jun 10 06
            Posts: 48
            Credit: 6,120,995
            RAC: 4,035
            Message 113827 - Posted 2 Sep 2011 12:23:43 UTC - in response to Message 111687.

              Depending on which GPU you add to your system, you may be able to run multiple work units at once per GPU using an app_info.xml file. Specifically the 4xx and 5xx Fermi GPUs can run two or more work units at a time optimally provided there is enough GPU memory. Each work units requires around 300-350 MB of GPU memory.


              Can you be more specific about the app_info.xml file? Where do I find it and what do I have to put in it?
              ____________

              Team Belgium

              Richard Haselgrove
              Send message
              Joined: Dec 10 05
              Posts: 1394
              Credit: 32,988,906
              RAC: 48,055
              Message 113829 - Posted 2 Sep 2011 13:33:13 UTC - in response to Message 113827.

                I'd recommend that you start with the driver stuff on it's own - you don't need an app_info.xml file for that.

                Get it running, check some tasks through and make sure that they validate. The time for moving up to multiple copies is after you've got the plain vanilla version running OK.

                Profile microchip
                Send message
                Joined: Jun 10 06
                Posts: 48
                Credit: 6,120,995
                RAC: 4,035
                Message 113940 - Posted 8 Sep 2011 14:14:26 UTC

                  Guys, a final question...

                  Do I need a specific driver version or can I grab the 280.13 one?

                  I'm asking because on http://developer.nvidia.com/cuda-toolkit-40 I see there's an older driver version (270.41.19) while the normal driver download page of nVIDIA offers version 280.13... so which one should I get ?

                  Thanks
                  ____________

                  Team Belgium

                  Richard Haselgrove
                  Send message
                  Joined: Dec 10 05
                  Posts: 1394
                  Credit: 32,988,906
                  RAC: 48,055
                  Message 113941 - Posted 8 Sep 2011 14:32:30 UTC - in response to Message 113940.

                    You certainly don't need a developer toolkit. Just the ordinary 280.13 should do it, though we have been having some problems with downclocking with the newer *windows* drivers. If you seem to have any problems like that, look for a menu item for "Beta and Archived Drivers" on the general - not developer - download page, and you should be able to find what you need there.

                    Profile microchip
                    Send message
                    Joined: Jun 10 06
                    Posts: 48
                    Credit: 6,120,995
                    RAC: 4,035
                    Message 113943 - Posted 8 Sep 2011 15:35:41 UTC

                      OK... thanks Richard for your help :)
                      ____________

                      Team Belgium

                      Post to thread

                      Message boards : Cruncher's Corner : CUDA on Linux


                      Home · Your account · Message boards

                      This material is based upon work supported by the National Science Foundation (NSF) under Grants PHY-1104902, PHY-1104617 and PHY-1105572 and by the Max Planck Gesellschaft (MPG). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the investigators and do not necessarily reflect the views of the NSF or the MPG.

                      Copyright © 2014 Bruce Allen