6.6 Deploying GPU + MPI-programs
It is possible to utilize multiple GPUs to the benefit of better application performance. Normally this is done by creating an MPI-program and making sure each MPI-task is connected to a GPU.
When we have Hyper-Q/MPS activated on Taito-GPU system, each GPU can be shared between multiple MPI-tasks of the same host CPU.
The MPI message passing is usually done by sending and receiving messages between host CPUs. However, with the advent of new CUDA-aware MPI libraries (like OpenMPI & MVAPICH2) it is possible to exchange messages between GPU resident data, so being able to pass the MPI implementation a pointer to memory resident on the GPU.
This may be beneficial, since not only your code becomes more readable (e.g. OpenACC versions), but also you avoid all the hassle in updating back & forth the data on host CPU just for the sake of making the message passing to work.
When using multiple nodes one needs to be extra verbose in the batch script to ensure the correct placement of the tasks. The easiest way to ensure correct placement is to use the
--ntasks-per-node flag to specify how many task each node should be running. The
--gres flag should be set to how many GPUs per node the job will use.
For example to run 4 tasks on 2 nodes and use 2 GPUs per node the following batch script should be used:
#!/bin/bash #SBATCH -N 2 #SBATCH -n 4 #SBATCH -p gpu #SBATCH -t 00:05:00 #SBATCH -J gpu_job #SBATCH -o gpu_job.out.%j #SBATCH -e gpu_job.err.%j #SBATCH --gres=gpu:k80:2 #SBATCH --ntasks-per-node= 2 #SBATCH module purge module load gcc cuda module list srun ./your_binary