Struct MpsDaemonManager 

pub struct MpsDaemonManager { /* private fields */ }

Manages NVIDIA MPS daemon instances, one per physical GPU.

The MPS control daemon must run on the host for each GPU that uses MPS sharing. This manager starts the daemon on first MPS container creation and stops it when the last MPS container on that GPU exits.

Implementations

impl MpsDaemonManager

pub fn new(base_dir: impl Into<PathBuf>) -> Self

Create a new MPS daemon manager.

base_dir is the root directory for per-GPU MPS pipe and log directories (e.g. /var/run/zlayer/mps).

pub async fn acquire(&self, gpu_index: u32) -> Result<PathBuf, MpsError>

Start or attach to the MPS daemon for a GPU.

If the daemon is already running for this GPU, increments the reference count. Otherwise, starts a new nvidia-cuda-mps-control daemon process.

Returns the pipe directory path that should be injected as CUDA_MPS_PIPE_DIRECTORY into the container’s environment.

Errors

Returns an error if the MPS control daemon fails to start.

pub async fn release(&self, gpu_index: u32)

Release a reference to the MPS daemon for a GPU.

Decrements the reference count. When it reaches zero, the MPS daemon is shut down by sending “quit” to the control pipe.

pub fn env_vars(&self, gpu_index: u32) -> Vec<(String, String)>

Get the environment variables needed for a container using MPS on a specific GPU.

pub async fn enable_time_slicing(gpu_index: u32) -> Result<(), MpsError>

Configure time-slicing for a GPU via nvidia-smi.

This sets the GPU to shared compute mode, allowing multiple processes to use it with round-robin scheduling.

Errors

Returns an error if nvidia-smi fails.

Trait Implementations

impl Debug for MpsDaemonManager

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.