The effect of neither nv_peer_mem or nvidia peermem detected


I got a warming when I compile the code:
/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.2/main_nvshmem/src/comm/transports/ibrc/ibrc.cpp:nvshmemt_init:1747: neither nv_peer_mem, or nvidia_peermem detected. Skipping transport.

I run the code in 1 node(8 GPUs). I would like to know if this warming actually effect the the performance of nvshmem_float_p or nvshmemx_float_put_nbi_block to communicate with other GPU.

The code is from: multi-gpu-programming-models/nvshmem/ at master · NVIDIA/multi-gpu-programming-models · GitHub



This should be a runtime warning.

It’s only applicable if you are running internode. nv_peer_mem, nvidia_peermem, and dmabuf (not mentioned here) are all used for GPUDirect RDMA.

On a single node where all GPUs are P2P connected, either through PCIe or NVLink, they will not be used - you should see no functional or performance issues.

If you would like to remove this runtime warning, you can set NVSHMEM_REMOTE_TRANSPORT=NONE in the environment.