Slow MPS performance on Multi-GPU system

I am attempting to debug poor weak I am seeing using MPS on a multi-GPU node. I am seeing the following message in the control.log file:

User did not send valid credentials

This message shows up even when using a single GPU on the node. The only reference I have found to this issue is https://github.com/NVIDIA/nvidia-docker/issues/868, however they do not address the issue of invalid credentials.

Can anyone point me to a discussion of what this message means, and how I might be able to resolve it?

Thanks,
Jacob