CUDA™ is a parallel computing platform and programming model invented by NVIDIA. It enables dramatic increases in computing performance by harnessing the power of the graphics processing unit (GPU). With millions of CUDA-enabled GPUs sold to date, software developers, scientists and researchers are finding broad-ranging uses for GPU computing with CUDA.
Research Computing has a significant investment in GPU-enabled servers. Each supports from two to eight Nvidia GPUs that range from the earlier S1070s (Tesla T10) to the more recent Kepler K20s (see table below).
|GPU||Quantity||Host Quantity||Host Architecture||Host Memory||Host Interconnect||Host Attributes||Notes|
|S1070||8||4||Intel E5462||16 GB||DDR IB||tesla,t10|
|M2070||8||4||Intel E5675||24 GB||QDR IB||fermi,m2070|
|M2070||8||1||Intel E5620||24 GB||GigE||fermi,m2070|
|M2090||10||5||Intel E5620||64 GB||FDR IB||fermi,m2090|
|M2090||16||8||AMD Opteron 6220||32 GB||QDR IB||fermi,m2090|
|K20c||2||1||Intel E5675||24 GB||QDR IB||k20||Reserved|
If you need interactive access to a gpu for development and testing you may do so by requesting an interactive session through the batch system.
In order to gain interactive access to a GPU server you should run similar to the one that follows.
qsub -I -l nodes=1:gpus=1:tesla,walltime=01:00:00 -q gpu
To gain access to one of the Fermi-class GPUs, you can make a similar request but specify the "fermi" attribute in your resource request as below.
qsub -I -l nodes=1:gpus=1:fermi,walltime=01:00:00 -q gpu
If a gpu is available, you will get a prompt on a gpu-enabled host within a minute or two. Otherwise, you will have to wait or try another time. If you choose to wait, you will be connected when a gpu is available. The default walltime limit for the gpu queue is 10 minutes. You should request the amount of time you need but be sure to log out and end your session when you are finished so that the GPU will be available to others.
If you need two GPUs in a single host, you would run the following command instead.
qsub -I -l nodes=1:gpus=2,walltime=01:00:00 -q gpu
If you need two gpus in two separate hosts, you would request
qsub -I -l nodes=2:gpus=1,walltime=01:00:00 -q gpu
The process is much the same for batch jobs. To access a host with an M2090, you can add the following to your submission script.
#PBS -q gpu #PBS -l nodes=1:gpus=1:m2090 #PBS -l walltime=1:00:00
To access a server with an M2070 GPU, you can add the following to your submission script.
#PBS -q gpu #PBS -l nodes=1:gpus=1:m2070 #PBS -l walltime=1:00:00
The GPUs are configured to run in exclusive mode. This means that the gpu driver will only allow one process at a time to access the GPU. If GPU 0 is in use and your application tries to use it, it will simply block. If your application does not call cudaSetDevice(), the CUDA runtime should assign it to a free GPU. Since everyone will be accessing the GPUs through the batch system, there should be no over-subscription of the GPUs.
For CUDA development please load the "cuda" module. Doing so will ensure that your environment is set up correctly for the use of the CUDA compiler, header files, and libraries.
$ module spider cuda Rebuilding cache, please wait ... (not written to file) done Description: NVIDIA CUDA Toolkit Versions: cuda/4.2 cuda/5.5 $ module load cuda/5.5 $ which nvcc /opt/cuda/5.5/bin/nvcc $ printenv | grep CUDA HPC_CUDA_LIB=/opt/cuda/5.5/lib64 HPC_CUDA_DIR=/opt/cuda/5.5 HPC_CUDA_BIN=/opt/cuda/5.5/bin HPC_CUDA_INC=/opt/cuda/5.5/include
PBS Script Examples
See the CUDA_PBS page for cuda PBS script examples.
See the CUDA_Examples page for CUDA development examples.