pagelocked

Pagelocked memory, also called page-locked or pinned memory, refers to host (CPU) memory that the operating system locks into physical RAM, preventing its pages from being swapped out to disk. In GPU computing, pagelocked memory is used to accelerate data transfers between the host and the device by allowing the GPU to access the memory directly through DMA without incurring page faults on each transfer.

In CUDA, pagelocked memory is typically allocated with functions such as cudaMallocHost or cudaHostAlloc. This memory can be used for host-device transfers and, in some cases, can be mapped to device pointers for direct access by the GPU. Pinned memory enables asynchronous data transfers that can be overlapped with computation, improving overall throughput. It can also enable zero-copy access on devices with unified addressing, though zero-copy generally comes with lower bandwidth compared to standard device memory.

Advantages of pagelocked memory include higher transfer bandwidth, reduced latency due to avoided page faults, and the ability to overlap transfers with computation through asynchronous streams. These benefits make it a common choice for streaming data into or out of GPUs and for host-side preparations that require frequent transfers.

Caveats and limitations include increased memory pressure on the host, since the OS cannot page this memory out. Locking large amounts of memory can lead to fragmentation, allocation failures, or reduced performance for other processes. Pinned memory should be used judiciously and typically in portions sized to the application's data transfer needs rather than locking all available RAM.