Very slow page allocation

[attila_afra]

Memory allocation is very slow on the device because of the expensive page faults. I've measured a page fault handling speed of about 1 GB/s. This means that the allocation/registration of large buffers can take several seconds! Are there any plans to improve the efficiency of page allocation in the kernel? I think this is a quite significant performance issue.

[TaylorIoTKidd]

Yes, this is a significant performance issue. There are some optimizations in the MPSS 3.2 release scheduled for the later part of this quarter (Q1 2014), and there are more to come in further releases.

One of our experts suggests that you look at using mmap() with the MAP_POPULATE flag to force the pages to be prepopulated, avoiding later page faults.

Regards
--
Taylor
 

[Patrick_S_]

try the code sample from intel for mmap

in the attachment you can find the code.