Scheduling GPU resources in the Grid

{i} Full documentation at this page.

Underlying tale of installations, applied configuration and tweaks

NVIDIA worker

  1. Blacklist nouveau.

    • To avoid compilation errors (aka ERROR: Unable to load the kernel module 'nvidia.ko'...) when installing NVIDIA driver, is often not enough to include blacklist nouveau in /etc/modprobe.d/blacklist.conf. It is also required to remove it from the initrd image like so: 

          # echo -e "blacklist nouveau\noptions nouveau modeset=0" > /etc/modprobe.d/disable-nouveau.conf
    # mkinitrd -f -v /boot/initrd-$(uname -r).img $(uname -r)  # or `dracut -f`
      and reboot.

  2. Install ./NVIDIA-Linux-x86_64-325.08.run or whatever other version.

  3. Install ./cudatoolkit_4.0.17_linux_64_rhel6.0.run or whatever other version.

  4. Check nvidia-smi command output.

    • If not supported or NA information is found: 
          (..)
    +-----------------------------------------------------------------------------+
    | Compute processes:                                               GPU Memory |
    |  GPU       PID  Process name                                     Usage      |
    |=============================================================================|
    |    0            Not Supported                                               |
    |    1            Not Supported                                               |
    |    2            Not Supported                                               |
    |    3            Not Supported                                               |
    +-----------------------------------------------------------------------------+

      we need to patch libnvidia-ml.so.1 library:

      1. Get patch from Github's nvml_fix repository.

      2. Compile it with TARGET=<your-nvidia-driver-version> (must be supported by the fix).

        • HACK: in Scientific Linux 6 it must be compiled with pthread and dl libraries: 

          •            # cat Makefile
           (..)
           CFLAGS        = -lpthread -ldl
           (..)

      3. Remove the link /usr/lib64/libnvidia-ml.so.1 and substitute it with the just created $PWD/libnvidia-ml.so.1 file.

        • Note that we use lib64 (not the default Makefile's libdir location -> lib).

        • Do not use make install PREFIX=/usr, copy it by hand.

        • Do not create a link, since ldconfig will overwrite it.

      Now nvidia-smi output should look like: 

          (..)
    +-----------------------------------------------------------------------------+
    | Compute processes:                                               GPU Memory |
    |  GPU       PID  Process name                                     Usage      |
    |=============================================================================|
    |  No running compute processes found                                         |
    +-----------------------------------------------------------------------------+

CREAM CE

  1. Added to BLAHP script /usr/libexec/sge_local_submit_attributes.sh: 

    •     (..)
    if [ -n $gpu ]; then
        echo "#$ -l gpu=${gpu}"
    fi
    (..)

Scheduler

  1. [qmaster] Define complex value 'gpu': 
    •     #name               shortcut     type        relop requestable consumable default  urgency  
    #-------------------------------------------------------------------------------------------
    (..)
    gpu                 gpu          INT         <=    YES         YES        0        0
    (..)
  2. [qmaster] Host(s) complexes: 
    •     hostname              tesla.ifca.es
    load_scaling          NONE
    complex_values        gpu=4,mem_free=24G,virtual_free=24G
    user_lists            NONE
    xuser_lists           NONE
    projects              NONE
    xprojects             NONE
    usage_scaling         NONE
    report_variables      NONE
  3. Load sensor: 
    • hostname=`uname -n`

while [ 1 ]; do
  read input
  result=$?
  if [ $result != 0 ]; then
    exit 1
  fi
  if [ "$input" == "quit" ]; then
    exit 0
  fi


  smitool=`which nvidia-smi`
  result=$?
  if [ $result != 0 ]; then
    gpusav=0
    gpus=0
  else
    gpustotal=`nvidia-smi -L|wc -l`
    gpusused=`nvidia-smi |grep "Process name" -A 6|grep -v +-|grep -v \|=|grep -v Usage|grep -v "No running"|wc -l`
    gpusavail=`echo $gpustotal-$gpusused|bc`
  fi

  echo begin
  echo "$hostname:gpu:$gpusavail"
  echo end
done

exit 0
  4. [qmaster] Per-host load sensor: 
    •     # qconf -sconf tesla
    #tesla.ifca.es:
    load_sensor                  /nfs4/opt/gridengine/util/resources/loadsensors/gpu.sh
    • Must be available in the execution node (e.g. shared via NFS)
  5. [execd] Restart execd process to load the new sensor: 
    • # ps auxf
(..)
root     24786  0.0  0.0 163252  2268 ?        Sl   16:51   0:00 /nfs4/opt/gridengine/bin/lx-amd64/sge_execd
root     24798  0.0  0.0 106104  1260 ?        S    16:51   0:00  \_ /bin/sh /nfs4/opt/gridengine/util/resources/loadsensors/gpu.sh
root     24801  0.0  0.0 106104   544 ?        S    16:51   0:00      \_ /bin/sh /nfs4/opt/gridengine/util/resources/loadsensors/gpu.sh
root     24802 71.0  0.0  11140   988 ?        R    16:51   0:00          \_ nvidia-smi -L
root     24803  0.0  0.0 100924   632 ?        S    16:51   0:00          \_ wc -l
(..)
    • soft-stop the service if there are jobs running.

  6. [qmaster] Query the GPU-host gpu resource: 

    •    # qhost -h tesla -F gpu
   HOSTNAME                ARCH         NCPU NSOC NCOR NTHR  LOAD  MEMTOT  MEMUSE  SWAPTO  SWAPUS
   ----------------------------------------------------------------------------------------------
   global                  -               -    -    -    -     -       -       -       -       -
   tesla                   lx-amd64        4    1    4    4  0.19   23.5G    1.7G   11.8G     0.0
       Host Resource(s):      hl:gpu=4.000000

References

  1. GridEngine

  2. NVIDIA CUDA