does dcgm on the physical host support capturing gpu metrics of kvm virtual machines?

[lddlww]

when there are many virtual machines installed on the physical host machine. These virtual machines use GPU through graphics card pass-through . I don’t want to install dcgm-exporter on each virtual machine; if I install dcgm-exporter on the host machine, can it capture all GPU metrics of virtual machine? if not，how can i do to solve it？

best wishes

[nikkon-dev]

@lddlww,

In passthrough mode, the GPU is not visible to the host OS. Therefore, the nv-hostengine (the server-side component of DCGM) must be run within the VM to which the GPU is attached.

The nv-hostengine can be configured to listen on a TCP port, allowing the DCGM exporter on the host to connect to it (provided the TCP port is exposed from the VM). However, this approach has limitations: the DCGM exporter does not currently support connecting to multiple nv-hostengine instances.