nvidia_p2p_get_pages() failed about gdrcopy HOT 7 CLOSED

Turns out, my driver and CUDA combination were not compatible.

from gdrcopy.

Hi @mdclemen,

I am also facing the same issue, but I couldn't find which the driver and CUDA version incur the error. I am wondering you did run gdrcopy with the lower version of driver or CUDA than the version specified in README.md at that time?

In my case, I am using following driver and CUDA software on two RTX2080-ti GPUs.

• Driver Version: 430.14
• CUDA Version: 10.2

If you have any suggestion, it would be great.

Thanks,
Jeongseob

from gdrcopy.

Hi @drossetti,

I do not have permission to reopen this issue. So, I just wanted you in this loop. Can you have any suggestions to resolve this issue, failing nvidia_p2p_get_pages() while running the validate binary?

from gdrcopy.

@Jeongseob unfortunately your GeForce card does not support GPUDirect RDMA feature. Please read carefully the prerequisites section of the README.md file.

from gdrcopy.

Ah, I just assumed that RTX2080 (Turing architecture) supports GPUDirect RDMA, but it is not.

Thanks!

from gdrcopy.

@drossetti is there any device property query that indicates a card supports RDMA or not?

from gdrcopy.

for the records, sometimes during CUDA 11.6, the attribute below was added:
CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED: Device supports GPUDirect RDMA APIs, like nvidia_p2p_get_pages (see https://docs.nvidia.com/cuda/gpudirect-rdma for more information)

from gdrcopy.