Quick Start

For an in-depth guide on MPS usage, refer to Appendix: Common Tasks.

For an in-depth guide on commands, environment variables, and more, refer to Appendix: Tools and Interface Reference.

Starting the MPS controller

To start the MPS controller as a daemon (recommended):

nvidia-cuda-mps-control -d

Launching an application under MPS

When the MPS controller is active, CUDA applications will use MPS:

./cuda_application

This can be verified with the ps command while the application is running:

echo ps | nvidia-cuda-mps-control

This can also be verified by using nvidia-smi after the application is launched to verify the existence of the nvidia-cuda-mps-server process. While the application is running, it will also appear in nvidia-smi with the M+C type specified.

Quitting MPS

To quit the MPS controller and any associated MPS servers:

echo quit | nvidia-cuda-mps-control

Explicitly starting an MPS server

To explicitly start an MPS server for user $UID:

echo start_server -uid $UID | nvidia-cuda-mps-control

Note that servers are started implicitly when a CUDA application is launched while the MPS controller is active.

Starting multiple control/server pairs

To start multiple MPS servers for the same user, you can use different CUDA MPS directories to start different controllers.

export CUDA_MPS_PIPE_DIRECTORY=/tmp/nvidia-mps-0 # Select a location that's accessible to the given $UID
export CUDA_MPS_LOG_DIRECTORY=/tmp/nvidia-log-0 # Select a location that's accessible to the given $UID
nvidia-cuda-mps-control -d # Start an MPS controller on pipe 0

export CUDA_MPS_PIPE_DIRECTORY=/tmp/nvidia-mps-1 # Select a location that's accessible to the given $UID
export CUDA_MPS_LOG_DIRECTORY=/tmp/nvidia-log-1 # Select a location that's accessible to the given $UID
nvidia-cuda-mps-control -d # Start an MPS controller on pipe 1

To start an application under one of the controllers, the same CUDA_MPS_PIPE_DIRECTORY must be set for the controller and the application.

CUDA_MPS_PIPE_DIRECTORY=/tmp/nvidia-mps-0 CUDA_MPS_LOG_DIRECTORY=/tmp/nvidia-log-0 ./cuda_application # This will use the MPS controller on pipe 0
CUDA_MPS_PIPE_DIRECTORY=/tmp/nvidia-mps-1 CUDA_MPS_LOG_DIRECTORY=/tmp/nvidia-log-1 ./cuda_application # This will use the MPS controller on pipe 1

Starting an MPS server with MLOPart

To start a server configured to use MLOPart on supported devices for user $UID:

echo start_server -uid $UID -mlopart | nvidia-cuda-mps-control

For more information, refer to Memory Locality Optimized Partitions.

Enabling static SM partitioning

Static SM partitioning must be enabled at MPS controller start time:

nvidia-cuda-mps-control -d -S

For more information, refer to Static SM Partitioning.