Page Locked Memory

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey!

I have some questions about page-locked memory:

  1. page locked memory means, paging is not used, so it’s mapped to physical memory of the host. Is that right?

  2. Why is only page locked memory able to perform asynchronous datatransfers?

Thanks in advance!

2

  1. Right

  2. Because transfer is done by the GPU and chipset, CPU is not used at all. Only CPU knows how to fetch a page from swap.

3

Thank you very much! But I have some more Questions:

  1. Are there any disadvantages by using “mapped pinned memory” in opposite to “normal” memory?

  2. Is it possible, that using “normal” memory is faster in some cases? If yes, in which?

best regards

4

Take a look here: http://forums.nvidia.com/index.php?showtopic=98502

5

basicly

  1. Because RAM is limited, see #2. Also you need to do additional coding for that :-)

  2. Very rare, if you have too little RAM for your App, so that swapping is good.

6

If your data still resides in cache (say after preprocessing) then it can be faster for small/tiny chunks below 32k. But then again, cache-memory isn’t really normal either, is it?