Struct NumaWorkerPool 

pub struct NumaWorkerPool { /* private fields */ }

Pool of NUMA workers, one per node.

This pool manages dedicated worker threads that are pinned to specific NUMA nodes. When you request an allocation for a GPU, the pool automatically determines the GPU’s NUMA node and routes the request to the appropriate worker.

Implementations

impl NumaWorkerPool

pub fn global() -> &'static Self

Get the global worker pool.

The pool is created lazily on first access and lives for the entire process lifetime.

pub fn allocate_pinned_for_gpu( &self, size: usize, gpu_id: u32, ) -> Result<*mut u8, String>

Allocate CUDA pinned memory for a specific GPU (auto-detects NUMA node).

This method:

1. Determines the GPU’s NUMA node via nvidia-smi
2. Routes the allocation to a worker pinned to that node
3. The worker allocates and touches pages to ensure first-touch placement

Arguments

• size - Number of bytes to allocate
• gpu_id - CUDA device ID

Returns

Raw pointer to the allocated memory. Caller is responsible for freeing via cudarc::driver::result::free_host.