Gpus

We have 7 GPUs on the cluster. (June 2016). There is a separate queue gpu.q to manage jobs

To log in interactively to the gpu queue:

qlogin -q gpu.q

Each gpu is a GeForce GTX 980

/sbin/lspci | grep -i nvidia

Instructions for getting setup for GPU computation: http://docs.nvidia.com/cuda/cuda-getting-started-guide-for-linux/

NVidia drivers are installed in /usr/loca/cuda*. To use the 7.5 drivers, make sure these environment variables are set:

export PATH=/usr/local/cuda-7.5/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-7.0/lib64:$LD_LIBRARY_PATH

Check that the drivers are installed:

cat /proc/driver/nvidia/version

which should return

NVRM version: NVIDIA UNIX x86_64 Kernel Module  352.79  Wed Jan 13 16:17:53 PST 2016
GCC version:  gcc version 4.4.7 20120313 (Red Hat 4.4.7-16) (GCC)

Try compiling and running the sample programs:

mkdir -p /scratch/$USER/cuda-7.5_samples
cp -r /usr/local/cuda-7.5/samples /scratch/$USER/cuda-7.5_samples
cd /scratch/$USER/cuda-7.5_samples/
make

Run the sample program

/nfs/ge/bin/on-one-gpu - /scratch/$USER/cuda-7.5_samples/bin/x86_64/linux/release/deviceQuery

Here is a sample script to run amber:

/nfs/work/tbalius/MOR/run_amber/run.pmemd_cuda_wraper.csh

Here is an excerpt from script

##########
cat << EOF > qsub.amber.csh
#\$ -S /bin/csh
#\$ -cwd
#\$ -q gpu.q
#\$ -o stdout
#\$ -e stderr

# export CUDA_VISIBLE_DEVICES="0,1,2,3" 
# setenv CUDA_VISIBLE_DEVICES "0,1,2,3"
setenv AMBERHOME /nfs/soft/amber/amber14/ 
set amberexe = "/nfs/ge/bin/on-one-gpu - \$AMBERHOME/bin/pmemd.cuda"
##########

Note that we run the executable with on-one-gpu. This manages which gpus are used.

If you generate significant output, which is generally but not always true, it is important to write locally to scratch and then copy things over the network onto the disk. If you write large amounts of data directly to the NFS disk it can cause problems for others.

When encountering kernel errors for Nvidia

The specific error will be an API error such as this which can be viewed with dmesg:

dmesg | grep "NVRM":
API mismatch: the client has the version 352.93, but
NVRM: this kernel module has the version 352.79.  Please
NVRM: make sure that this kernel module and all NVIDIA driver
NVRM: components have the same version.

Address this problem with the following commands:

# this will remove the nvidia modules and reload the updated versions of the modules

rmmod nvidia_uvm
rmmod nvidia_drm
rmmod nvidia_modeset
rmmod nvidia
modprobe nvidia
modprobe nvidia_modeset
modprobe nvidia_uvm

monitor throttling

nvidia-smi

commands appear to be in /opt/set-gpu-fans/

Update CUDA 11 and Nvidia-driver

$ yum -y install yum-utils
$ yum-config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
$ yum-config-manager --disable repository cuda # disable existing cuda 10 repo
$ yum remove xorg-x11-drv-nvidia* -y
$ yum install nvidia-driver-latest-dkms cuda -y 

Troubleshooting I ran into this error while trying to 'yum install nvidia-driver-latest-dkms cuda -y'

Error: Package: 3:nvidia-driver-latest-dkms-cuda-libs-515.43.04-1.el7.x86_64 (cuda-rhel7-x86_64)
          Requires: libnvidia-egl-wayland.so.1()(64bit)

So I google the url link for this package and install it and proceed as usual

$ rpm -Uvh https://download-ib01.fedoraproject.org/pub/epel/7/x86_64/Packages/e/egl-wayland-1.1.6-1.el7.x86_64.rpm

For Failed to initialize NVML: Driver/library version mismatch error, follow the instruction above to reload the nvidia modules to match the kernel version