Allocating pinned memory with large RAM configurations

Accelerated Computing CUDA CUDA Programming and Performance
,
1

Hi all,

I’ve recently upgraded a workstation to 1TB of RAM and since doing so I have been unable to allocate any page locked memory of size >2GB. It works perfectly fine with any other RAM configuration (eg 768GB) and I know my PC supports the new 1TB config. I’m a bit lost over what could be causing this

I see the issue in pycuda calling cuda.pagelocked_empty(array_size), but also with a toolbox ( CERN/TIGRE: TIGRE: Tomographic Iterative GPU-based Reconstruction Toolbox) that allocates the memory using C++ and cudaHostRegister.

Is there any known issue that might be stopping page-locked memory allocation with large RAM configurations?

Thanks

2

For each allocation or 2GB in sum?

You could allocate several 1GB blocks as workaround?

3

If the issue is reproduceable with a minimal plain CUDA program, with the latest CUDA version (12.6 update 3) and driver package available, I would suggest filing a bug with NVIDIA.

There could be a bug of the “internal integer overflow” kind at the core of the problem. It could be due something in the CUDA software stack or part of the operating system. My assumption is that machines with >= 1TB of system memory are still sufficiently rare that the possibility of a software test escape is a possibility.

4

This may shed some light on the situation.