Is this page helpful?

MPS v3 Namespaces#

A namespace subdivides a server’s resources – pinned memory limit, active thread percentage, and client priority – for a group of clients. Servers are typically shared across many clients with different resource needs; namespaces let those groups of clients be managed independently without requiring a separate server per group.

Every server has an implicit default namespace, which is used by clients that do not target a specific namespace. Refer to MPS v3 Interface for the namespace create/delete/list/set/get command reference, and to MPS v3 Configuration File for configuring namespaces from a TOML file.

Namespace Routing#

A client selects its namespace by which socket, under the server’s pipe directory, it connects to:

$CUDA_MPS_PIPE_DIRECTORY/
├── control                          # main daemon control socket
├── training/                        # server "training"
│   ├── control                      # server listener socket (routes to the default namespace)
│   ├── default/
│   │   └── control                  # default namespace socket
│   ├── high_priority/               # custom namespace
│   │   └── control                  # namespace socket
│   └── batch/
│       └── control

  • $CUDA_MPS_PIPE_DIRECTORY/training/high_priority/control – routes to the high_priority namespace on server training.

  • $CUDA_MPS_PIPE_DIRECTORY/training/control – routes to server training’s default namespace.

  • $CUDA_MPS_PIPE_DIRECTORY/control – UID-based auto-routing, for backward compatibility with Legacy MPS v2 clients.

A client sets CUDA_MPS_PIPE_DIRECTORY to the namespace’s directory to connect through that namespace’s socket, in the same way Legacy MPS v2 clients set it to a server’s pipe directory.

On this page