NCCL allreduce hangs when cudaFreeHost about nccl

[fmana]

I forgot to add some info:

1] I'm using NCCL 1.1.1.
2] My application is multi-thread and each thread uses a GPU (so multi-GPU too)
3] When the application falls into deadlock (N-1) GPU load is 100%, 1 GPU load 0%.
Probably (N-1) GPU polling to shake-their-hand condition,
1 GPU into cuFreeHost (enqueued in someplace)

[CFAndy]

Hi Fmana, any update?

[fmana]

The issue still remain and to overcome it I need to use CPU-based barrier in order to
protect NCCL allreduce primitive.
So I need to have fixed the CUDA programming model limitation as noticed in #37.

Any news on issue/question #49 about nccl deployment for future?

[sjeaugey]

Should be fixed since NCCL 2.1. Please re-open if still a problem.