0

I have some questions about pinned memory and paged memory

I saw some papers said that if we use pinned memory, we can get less latency of memory copy(Host to Device) than use paged memory

(e.g. Reducing GPU Offload Latency via Fine-Grained CPU-GPU Synchronization by Daniel Lustig and Margaret Martonos )

There are few questions as below:

  1. What makes pinned memory more faster than paged memory? (Because I still don't know the different between them)
  2. Is this only use on CUDA environments?
Hsu Chao Yi
  • 1
  • 1
  • PInned memory is based on feature of the PCI Express specification. It is supported in OpenCL and OpenGL as well, and I think other vendors of PCI-e discrete GPUs support a similar DMA based transfer mode – talonmies Jul 16 '15 at 05:18
  • @talonmie Thanks , so we can said because Pinned memory using DMA for memory copy, so it makes sense that will cost less latency than normal paged memory, but I am thinking that if program set pinned memory larger will lead the Cache-coherence problem ? – Hsu Chao Yi Jul 16 '15 at 05:35

0 Answers0