NVIDIA Multi-Process Service (MPS)

NVIDIA Multi-Process Service (MPS) is a runtime architecture that allows multiple GPU-enabled processes, typically MPI jobs, to run concurrently and share resources more efficiently on a single GPU. MPS is recommended for tasks that individually do not use all the compute power of the GPU or that only contain fine-grained GPU operations (small number of blocks per grid or threads per grid, see Introduction to CUDA for terminology). The steps below outline how to launch multiple MPI tasks on a single GPU simultaneously using MPS. For full details, see the NVIDIA Multi-Process Service documentation or the Vista User Guide: NVIDIA MPS.