GPU use on NeSI

• What is a GPU?
  A GPU is a co-processor to accelerate CPUs for general-purpose computing.
  
  
• What is the difference between CPU and GPU?
  GPUs contain more computing units than CPUs.
  
  
• How many GPUs does NeSI have?
  NeSI has 16 nvidia Tesla P100 GPU cards.
  mahuika has 8 GPU cards distributed over 4 GPU nodes (2 by node).
  maui ancil has 4 GPU cards distributed over 4 GPU nodes (1 per node).
  4 GPU cards are used by NIWA for VM.

• SLURM job for GPU.
  Add the option #SBATCH --gpus-per-node in a Slurm script.

#!/bin/bash -e
#SBATCH --job-name=GPUJob   #job name (shows up in the queue)
#SBATCH --time=00:00:01     #Walltime (HH:MM:SS)
#SBATCH --gpus-per-node=1   #GPU resources required per node
srun pwd # Prints working directory

• SLURM interactive session for GPU
  Add the option --gpus-per-node to the command salloc.

  @mahuika02]$ salloc --gpus-per-node=1
  salloc: Pending job allocation XXXXX
  salloc: job XXXXX queued and waiting for resources
  salloc: Nodes wbg001 are ready for job
  @mahuika02]$ ssh wbg001
  @wbg001]$

See also: https://support.nesi.org.nz/hc/en-gb/articles/360001316356-Slurm-Interactive-Sessions

• Which is better for my code between CPU or GPU?
  Run benchmarks to measure speedup with CPUs and GPUs.
• Offload computation on the GPU
  Use OpenMP, OpenCL, OpenACC or CUDA directive to offload computation on the GPU.
• What is OpenMP?
  OpenMP is a library that supports shared memory multiprocessing.

• What is OpenCL or OpenACC?
  OpenCL or OpenACC are libraries that support accelerators like GPU.
  
  
• What is CUDA?
  CUDA is a library that supports nvidia GPUs.

• GPU monitoring
  Use nvidia-smi on the command line to monitor directly on the GPU.

• CUDA compilation.
   Use nvcc on the command line to compile a CUDA program.

$ ml CUDA
$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver

See also: https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html

• CUDA debugging.
  Use cuda-gdb on the command line to debug a CUDA applications on the login node (for mahuika ) or directly on the GPU (for mahuika and maui ancil).

$ ml CUDA
$ cuda-gdb --version
NVIDIA (R) CUDA Debugger

 See also: https://docs.nvidia.com/cuda/cuda-gdb/index.html

• CUDA profiling.
  Use Nsight compute on the command line to profile a CUDA applications directly on the GPU.

$ ml CUDA
$ nv-nsight-cu-cli --version
NVIDIA (R) Nsight Compute Command Line Profiler

See also: https://docs.nvidia.com/nsight-compute/2020.1/NsightComputeCli/index.html