The development version of Amber14 currently does not support P2P for
CUDA_VISIBLE_DEVICES="0,1" or
CUDA_VISIBLE_DEVICES="2,3".
it gives the error:
cudaDeviceEnablePeerAccess all CUDA-capable devices are busy or unavailable
Work-around:
To get around this issue, root must disable exclusive compute mode on the
relevant cards using "nvidia-smi -c 0"
[root.compute-0-1 ~]# nvidia-smi -c 0
Set compute mode to DEFAULT for GPU 0000:02:00.0.
Set compute mode to DEFAULT for GPU 0000:03:00.0.
Set compute mode to DEFAULT for GPU 0000:83:00.0.
Set compute mode to DEFAULT for GPU 0000:84:00.0.
And now....
---------------- GPU PEER TO PEER INFO -----------------
|
| Peer to Peer support: ENABLED
|
|--------------------------------------------------------
Ross has submitted this bug to NVIDIA, but this work-around will suffice in
the interim.
Happy simulating!
--Chris McClendon, UCSD
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat Mar 08 2014 - 18:00:03 PST
