Problems when using MPS

Hi, I am trying to use Multi-process service to test its performance.
I configured my environment as the steps shown in https://docs.nvidia.com/deploy/pdf/CUDA_Multi_Process_Service_Overview.pdf .

And I run two processes concurrently.

However, when I read the control.log, It shows like this:
[2017-09-25 13:46:47.050 Control 31085] Accepting connection…
[2017-09-25 13:46:47.050 Control 31085] User did not send valid credentials
[2017-09-25 13:46:47.050 Control 31085] Accepting connection…
[2017-09-25 13:46:47.050 Control 31085] NEW CLIENT 7547 from user 1000: Server already exists
[2017-09-25 13:49:23.785 Control 31085] Accepting connection…
[2017-09-25 13:49:23.785 Control 31085] User did not send valid credentials
[2017-09-25 13:49:23.785 Control 31085] Accepting connection…
[2017-09-25 13:49:23.786 Control 31085] NEW CLIENT 9484 from user 1000: Server already exists

So, what’s the meaning of “User did not send valid credentials”? Could anyone please give me some advice?

Thanks a lot!

@xiaodongyee I’m facing similar problem. Did you find any solution to this?

We are also experiencing this!

Feedback would be very appreciated

(nbody MPS Samples on DGX-1)

I’m also experiencing this. My system is a DELL with 8xK80,

Any help it is appreciated.

Regards,
JP

Does anyone has solution?

I notice this happen if the MPS client is inside docker machine.

I can start the docker with --ipc=host but it will not work inside ECS container

This may be of interest, regarding MPS inside docker:

[url]https://github.com/NVIDIA/nvidia-docker/wiki/MPS-(EXPERIMENTAL)[/url]

I am aware it is an old topic but it is still without solution. Any progress/solution for this issue? I also have the error “User did not send valid credentials” in /var/log/nvidia-mps/control.log on bare metal. Haven’t tried mps in a docker container.