Discussion Thread for the Continuous GW Search known as O2MD1 (now O2MDF - GPUs only)

Forums › Cruncher's Corner

Trotador
Trotador
Joined: 2 May 13
Posts: 58
Credit: 2122643213
RAC: 0

Vega 64 and RX70 seem to be

24 Nov 2019 19:32:17 UTC
Message 174543
(moderation:
    )

    Vega 64 and RX70 seem to be crunching Ok at x4 wu at once but Radeon VII does not, GPU doesn't get loaded even with 6 wus, Could be my setup but FGRPB1G and milky crunch OK. All in Ubuntu.

    Gary Roberts
    Gary Roberts
    Moderator
    Joined: 9 Feb 05
    Posts: 5872
    Credit: 117693059168
    RAC: 35069735

    B.I.G wrote:... hope the devs

    24 Nov 2019 22:57:54 UTC
    Message 174546 in response to message 174542
    (moderation:
      )

      B.I.G wrote:
      ... hope the devs manage to fix this. I'd love to crunch GW tasks.

      Please be realistic :-).  How could the Devs fix your setup?  You need to identify what's wrong with your system.

      A good thing to start with would be to run the clinfo utility (if you don't have it installed, go find it) and see exactly what OpenCL platforms are reported.  There's lots of output but the platforms you have will be identified quite early in the output..  For example, here's what I see for an RX 570.

      Number of platforms:                             1
  Platform Profile:                              FULL_PROFILE
  Platform Version:                              OpenCL 2.1 AMD-APP (2671.3)
  Platform Name:                                 AMD Accelerated Parallel Processing
  Platform Vendor:                               Advanced Micro Devices, Inc.
  Platform Extensions:                           cl_khr_icd cl_amd_event_callback cl_amd_offline_devices

  Platform Name:                                 AMD Accelerated Parallel Processing
Number of devices:                               1
  Device Type:                                   CL_DEVICE_TYPE_GPU
  Vendor ID:                                     1002h
  Board name:                                    Radeon RX 570 Series
  Device Topology:                               PCI[ B#7, D#0, F#0 ]
  Max compute units:                             32

      It could well be that you might need a more recent version of OpenCL.  Notice mine says 2.1.  I'm pretty sure you would be best to ensure your version is at least 1.2.  I think there could well be stuff in later versions that might improve your performance.

      Let's just start by seeing what you actually have.  If it needs upgrading, that will be out of my league for something like MacOS.

      Cheers,
      Gary.

      B.I.G
      B.I.G
      Joined: 26 Oct 07
      Posts: 117
      Credit: 1176702118
      RAC: 990828

      Couldn't find clinfo utility

      24 Nov 2019 23:52:38 UTC
      Message 174547
      (moderation:
        )

        Couldn't find clinfo utility for macOS but according to Apple macOS is using Open CL 1.2 

        https://support.apple.com/en-us/HT202823

        And there is the Problem: I can't update Open CL, I can't fix the System it is as provided by Apple. Either Apple got it right or wrong. This is not Linux or Windows.

        Keith Myers
        Keith Myers
        Joined: 11 Feb 11
        Posts: 4964
        Credit: 18751224800
        RAC: 7100427

        I thought that Apple is

        25 Nov 2019 0:33:04 UTC
        Message 174550
        (moderation:
          )

          I thought that Apple is abandoning both OpenGL and OpenCL and going with their closed Metal API for all current and future MacOS releases.  Distributed computing is at an end on Apple hardware as far as I can see.

          https://appleinsider.com/articles/18/06/04/opengl-opencl-deprecated-in-favor-of-metal-2-in-macos-1014-mojave

          And the same goes for CUDA. https://setiathome.berkeley.edu/forum_thread.php?id=84805&postid=2020081

           

          Gary Roberts
          Gary Roberts
          Moderator
          Joined: 9 Feb 05
          Posts: 5872
          Credit: 117693059168
          RAC: 35069735

          B.I.G wrote:Couldn't find

          25 Nov 2019 4:19:50 UTC
          Message 174551 in response to message 174547
          (moderation:
            )

            B.I.G wrote:
            Couldn't find clinfo utility for macOS but according to Apple macOS is using Open CL 1.2.

            A quick google search reveals that OpenCL 1.2 (which includes the clinfo utility) is installed in High Sierra up to at least 10.13.4.  I also saw an actual Apple page which said that existing OpenCL apps would still run in 10.14.  How well such apps might run is quite another matter.

            You should not expect BOINC projects to be able to do anything about this.  You should complain directly to Apple - not that that tactic would be likely to make any difference.  I don't know the current status but there used to be a BOINC developer who tweaked stuff so that BOINC could run properly on Apple products.  Maybe someone over at BOINC might know if BOINC still properly detects and deals with all the different hardware in Apple systems, particularly hardware needing OpenCL support.

            Individual projects rely on BOINC to properly detect the hardware capabilities.  BOINC relies on the OS/drivers to get that information.  Anyone contributing to BOINC with MacOS in mind should know the true situation, which is why asking on the BOINC boards might get you better information.

            Did you even try opening a command prompt (terminal window) and issuing the clinfo command?

            Cheers,
            Gary.

            Logforme
            Logforme
            Joined: 13 Aug 10
            Posts: 332
            Credit: 1714373961
            RAC: 0

            I have a question about the

            25 Nov 2019 19:20:21 UTC
            Message 174558
            (moderation:
              )

              I have a question about the CPU usage of the O2MDF app. I tried searching to see if it's been answered before but I found nothing.

              On this machine that have 4 CPU cores I used to have "Use at most: 100% of the processors" and "GPU utilization factor" set to 1.0.

              With the FGRPB1G app this resulted in 1 GPU task and 3 CPU tasks running. Each app getting one core, everyone is getting resources.

              With the O2MDF app this resulted in 1 GPU task and 4 CPU tasks running. This would starve the GPU app of CPU resources and cause it to be very slow.

              So nowadays I have to set "Use at most: 95% of the processors" which results in 1 O2MDF task and 3 CPU tasks. The problem is that when a FGRPB1G task comes along (not often, the O2MDF tasks dominate) the machine will only run 1 FGRPB1G task and 2 CPU tasks.

              Is this working as intended?

              Have I missed something in the configuration?

              Jim1348
              Jim1348
              Joined: 19 Jan 06
              Posts: 463
              Credit: 257957147
              RAC: 0

              Logforme wrote:Is this

              25 Nov 2019 20:44:19 UTC
              Message 174559 in response to message 174558
              (moderation:
                )

                Logforme wrote:

                Is this working as intended?

                Have I missed something in the configuration?

                Yes, that is as intended.  It is because FGRPB1G reserves a whole CPU core when it is running:

                Einstein@Home    1.20 Gamma-ray pulsar binary search #1 on GPUs (FGRPopencl1K-nvidia)  1C + 1 NVIDIA GPU  

                Whereas O2MDF reserves only 0.9 of a CPU core.

                You can fix that by using an app_config.xml file to force the O2MDF to use a whole core also, but I have not done it in a while.  I am sure someone here can give you a sample.

                Richie
                Richie
                Joined: 7 Mar 14
                Posts: 656
                Credit: 1702989778
                RAC: 0

                If you want to restrict CPU

                25 Nov 2019 21:07:55 UTC
                Message 174561
                (moderation:
                  )

                  If you want to restrict CPU tasks only to FGRP5 then here's one kind of example that might work for you:

                  <app_config>
                  <app>
                  <name>hsgamma_FGRP5</name>
                  <max_concurrent>3</max_concurrent>
                  </app>
                  <app>
                  <name>hsgamma_FGRPB1G</name>
                  <max_concurrent>1</max_concurrent>
                  <gpu_versions>
                  <gpu_usage>1</gpu_usage>
                  <cpu_usage>1</cpu_usage>
                  </gpu_versions>
                  </app>
                  <app>
                  <name>einstein_O2MDF</name>
                  <max_concurrent>1</max_concurrent>
                  <gpu_versions>
                  <gpu_usage>1</gpu_usage>
                  <cpu_usage>1</cpu_usage>
                  </gpu_versions>
                  </app>
                  </app_config>

                  Logforme
                  Logforme
                  Joined: 13 Aug 10
                  Posts: 332
                  Credit: 1714373961
                  RAC: 0

                  Jim1348 wrote:Yes, that is as

                  25 Nov 2019 22:05:27 UTC
                  Message 174562 in response to message 174559
                  (moderation:
                    )

                    Jim1348 wrote:

                    Yes, that is as intended.  It is because FGRPB1G reserves a whole CPU core when it is running:

                    Whereas O2MDF reserves only 0.9 of a CPU core.

                    Yes I did notice the 1.0 and 0.9 values.

                    I just find it strange that with 3 CPU tasks and 1 O2MDF task the math becomes:

                    3 * 1.0 + 1 * 0.9 = 3.9 ... Hmm .. we have 4.0 so lets start another 1.0 task. But I guess that's a Boinc issue, not E@H.

                    What's the rationale behind the 0.9 value for O2MDF? Why not 1.0?

                    Richie
                    Richie
                    Joined: 7 Mar 14
                    Posts: 656
                    Credit: 1702989778
                    RAC: 0

                    Richie wrote:Nvidia...I

                    26 Nov 2019 9:19:24 UTC
                    Message 174566 in response to message 174508
                    (moderation:
                      )

                      Richie wrote:
                      Nvidia...I recall 700-series doesn't work... everything from Kepler is incompatible ?

                      I saw a host with GT 630 2GB (Fermi chip, GF108) and it had successful validations under Windows.

                      https://einsteinathome.org/host/11687949

                      Outcome:Success

                      Validation state:Valid

                      OpenCL Device used for Search/Recalc and/or semi coherent step: 'GeForce GT 630 (Platform: NVIDIA CUDA, global memory: 2048 MiB)

                      Fermi is even older than Kepler. I would assume now that at least some of the 600- and 700-series cards are compatible. GT 650M was reported to work under Mac and it has Kepler GK107 indeed. But six different chips have been used in Kepler cards and desktop models up from GTX 660 were different in that way to GT 650M.

                      Some kind of lower bound for Nvidia compatibility may lie somewhere in Fermi architecture.

                      https://www.techpowerup.com/gpu-specs/?architecture=Fermi&sort=generation

                       

                      Quote:
                      But how about GTX 750 (Ti) ? It has Maxwell 'first generation' chip.

                      Somebody had been running these tasks with GTX 750 Ti and it worked well.

                      Comment viewing options

                      Select your preferred way to display the comments and click "Save settings" to activate your changes.

                      Forums › Cruncher's Corner