Fork me on GitHub

CUDA

CUDA (Compute Unified Device Architecture) is a parallel computing platform and application programming interface (API) model created by NVIDIA. It allows software developers to use a CUDA-enabled graphics processing unit (GPU) for general purpose processing, an approach known as General Purpose GPU (GPGPU) computing.

Usage

You need to first request one or more GPUs within an interactive session or batch job on a worker node. For example, to request a single GPU for an interactive session on a worker node:

qrshx -l gpu=1

Note

See Using GPUs on ShARC for more information on how to request a GPU-enabled node for an interactive session or job submission.

You then need to load a version of the CUDA library (and compiler). There are several versions of the CUDA library available. As with much software installed on the cluster, versions of CUDA are activated via the ‘module load’ command:

module load libs/CUDA/11.0.2/binary
module load libs/CUDA/10.2.89/binary
module load libs/CUDA/10.1.243/binary
module load libs/CUDA/10.0.130/binary
module load libs/CUDA/9.1.85/binary
module load libs/CUDA/9.0.176/binary
module load libs/CUDA/8.0.44/binary
module load libs/CUDA/7.5.18/binary

To then confirm which version of CUDA you are using:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2020 NVIDIA Corporation
Built on Thu_Jun_11_22:26:38_PDT_2020
Cuda compilation tools, release 11.0, V11.0.194
Build cuda_11.0_bu.TC445_37.28540450_0

Important To compile CUDA programs you also need a compatible version of the GCC compiler:

  • CUDA 7.x and 8.x: GCC >= 4.7.0 (to allow for the use of c++11 features) and < 5.0.0

  • CUDA 9.x: GCC < 7.0.0

Compiling a simple CUDA program

An example of the use of nvcc (the CUDA compiler):

nvcc filename.cu

will compile the CUDA program contained in the file filename.cu.

Compiling the sample programs

You do not need to be using a GPU-enabled node to compile the sample programs but you do need a GPU to run them.

In a qrshx session:

# Load modules
module load libs/CUDA/11.0.2/binary

# Copy CUDA samples to a local directory
# It will create a directory called NVIDIA_CUDA-11.0_Samples/
mkdir cuda_samples
cd cuda_samples
cp -r $CUDA_SDK .

# Compile (this will take a while)
cd NVIDIA_CUDA-11.0_Samples/
make

The make command then runs the nvcc CUDA compiler and generates a binary executable that you can then run on a node with an NVIDIA GPU installed.

A basic test is to run one of the resulting binaries, deviceQuery.

GPU Code Generation Options

To achieve the best possible performance whilst being portable, GPU code should be generated for the architecture(s) it will be executed upon.

This is controlled by specifying -gencode arguments to NVCC which, unlike the -arch and -code arguments, allows for ‘fatbinary’ executables that are optimised for multiple device architectures.

Each -gencode argument requires two values, the virtual architecture and real architecture, for use in NVCC’s two-stage compilation. I.e. -gencode=arch=compute_60,code=sm_60 specifies a virtual architecture of compute_60 and real architecture sm_60.

To support future hardware of higher compute capability, an additional -gencode argument can be used to enable Just in Time (JIT) compilation of embedded intermediate PTX code. This argument should use the highest virtual architecture specified in other gencode arguments for both the arch and code i.e. -gencode=arch=compute_60,code=compute_60.

The minimum specified virtual architecture must be less than or equal to the Compute Capability of the GPU used to execute the code.

Public GPU nodes in ShARC contain Tesla K80 GPUs, which are compute capability 37. To build a CUDA application which targets the public GPUS nodes, use the following -gencode arguments:

nvcc filename.cu \
   -gencode=arch=compute_37,code=sm_37 \
   -gencode=arch=compute_37,code=compute_37

ShARC also contains Tesla P100 GPUs and Tesla V100 GPUs in private nodes, which are compute capability 60 and 70 respectively. To build a CUDA application which targets any GPU on ShARC (either public or private), use the following -gencode arguments (for CUDA 9.0 and above):

nvcc filename.cu \
   -gencode=arch=compute_37,code=sm_37 \
   -gencode=arch=compute_60,code=sm_60 \
   -gencode=arch=compute_70,code=sm_70 \
   -gencode=arch=compute_70,code=compute_70

Further details of these compiler flags can be found in the NVCC Documentation, along with details of the supported virtual architectures and real architectures.

Note

SM 60 for Pascal GPUs is only available for CUDA 8.0 and above.

Note

SM 70 for Volta GPUs is only available for CUDA 9.0 and above.

Profiling using nvprof

Note that nvprof, NVIDIA’s CUDA profiler, cannot write output to the /fastdata filesystem.

This is because the profiler’s output is a SQLite database and SQLite requires a filesystem that supports file locking but file locking is not enabled on the (Lustre) filesystem mounted on /fastdata (for performance reasons).

CUDA Training

GPUComputing@sheffield provides a self-paced introduction to CUDA training course.

Determining the NVIDIA Driver version

Run the command:

cat /proc/driver/nvidia/version

Example output is:

NVRM version: NVIDIA UNIX x86_64 Kernel Module  450.51.05  Sun Jun 28 10:33:40 UTC 2020
GCC version:  gcc version 4.8.5 20150623 (Red Hat 4.8.5-39) (GCC)

Installation notes

These are primarily for system administrators.

Device driver

The NVIDIA device driver is installed and configured using the gpu-nvidia-driver systemd service (managed by Puppet). This service runs /usr/local/scripts/gpu-nvidia-driver.sh at boot time to:

  • Check the device driver version and uninstall it then reinstall the target version if required;

  • Load the nvidia kernel module;

  • Create several device nodes in /dev/.

CUDA 11.0.2

  1. Installed with install.sh with 11.0.2_450.51.05 as the sole argument.

  2. Modulefile was installed as /usr/local/modulefiles/libs/CUDA/11.0.2/binary

CUDA 10.2.89

  1. Installed with install.sh with 10.2.89_440.33.01 as the sole argument.

  2. Modulefile was installed as /usr/local/modulefiles/libs/CUDA/10.2.89/binary

CUDA 10.1.243

  1. Installed with install.sh with 10.1.243_418.87.00 as the sole argument.

  2. Modulefile was installed as /usr/local/modulefiles/libs/CUDA/10.1.243/binary

CUDA 10.0.130

  1. Installed with install.sh with 10.0.130_410.48 as the sole argument. This installs the toolkit and three NVIDIA-recommended patches.

  2. Modulefile was installed as /usr/local/modulefiles/libs/CUDA/10.0.130/binary

CUDA 9.1.85

  1. Installed with install.sh with 9.1.85_387.26 as the sole argument. This installs the toolkit and three NVIDIA-recommended patches.

  2. Modulefile was installed as /usr/local/modulefiles/libs/CUDA/9.1.85/binary

CUDA 9.0.176

  1. Installed with install.sh with 9.0.176_384.81 as the sole argument. This installs the toolkit and four NVIDIA-recommended patches.

  2. Modulefile was installed as /usr/local/modulefiles/libs/CUDA/9.0.176/binary

CUDA 8.0.44

  1. Installed with install.sh with 8.0.44 as the sole argument.

  2. This modulefile was installed as /usr/local/modulefiles/libs/CUDA/8.0.44/binary

CUDA 7.5.18

  1. Installed with install.sh with 7.5.18 as the sole argument.

  2. This modulefile was installed as /usr/local/modulefiles/libs/CUDA/7.5.18/binary