SOLVED Intel GPU Plugin

[patmage]

Is it possible to set the -shared-dev-num mentioned in https://intel.github.io/intel-devic.../gpu_plugin/README.html#configuration-options? I'd like to give 2 apps access to the GPU for transcoding and this document from Intel makes it sound like that is possible, but I don't know enough yet of kubernetes to know how to go about doing that.

 

[truecharts]

As far as we're aware you can already have multiple apps use the same GPU just fine. With both official Apps and TrueCharts.
Best advice is to first test it yourself first and report to the forum of something doesn't work. ;-)

 

[patmage]

Sorry, that is why I made this topic. I tried and the app will hang during deployment. "0/1 nodes are available: 1 Insufficient gpu.intel.com/i915 " is the error message I received. Looking at the Intel docs that I linked to the default is set to 1 container per GPU.

 

[ornias]

Shoot.
@waqarahmed You might want to fix this for 21.08 even.

 

[patmage]

For anyone else that finds this I think running "k3s kubectl edit ds/intel-gpu-plugin -n kube-system" and adding args: ["-shared-dev-num","5"] to the container section should work. At least now it is showing 5 available gpu resources.

 

[jsclayton]

I’d just stumbled on to the same issue trying to run Jellyfin alongside Plex. Got me wondering if the Intel GPU plugin should be deployed as an app rather than let the middleware handle it. This shared argument is the second issue I’ve had with it where the solution was to manually configure it in k8s.

The first was the middleware not recognizing the iGPU on a motherboard with IPMI. Until 21.06 I had to deploy the chart for the plug-in manually.

In both instances the way that the middleware applies the plug-in chart on boot wiped the manual changes (or removed it completely).

 

[ornias]

I’d just stumbled on to the same issue trying to run Jellyfin alongside Plex. Got me wondering if the Intel GPU plugin should be deployed as an app rather than let the middleware handle it. This shared argument is the second issue I’ve had with it where the solution was to manually configure it in k8s.

The first was the middleware not recognizing the iGPU on a motherboard with IPMI. Until 21.06 I had to deploy the chart for the plug-in manually.

In both instances the way that the middleware applies the plug-in chart on boot wiped the manual changes (or removed it completely).

Did you file a bug reports to get it fixed?

 

[ornias]

Submitted the PR to get it fixed quickly:

NAS-112058 / 21.10 / Allow intel GPU's to be used by multiple Apps by Ornias1993 · Pull Request #7431 · truenas/middleware

As discussed and not reported on Jira (afaik) here: https://www.truenas.com/community/threads/intel-gpu-plugin.94955/ This allows the Intel GPU's using the Intel GPU plugin, to be used by multiple ...

github.com

If this was submitted earlier it most likely woudl've been included in 21.08, now it's not-so-sure to be included.

 

[Rudi Pittman]

For anyone else that finds this I think running "k3s kubectl edit ds/intel-gpu-plugin -n kube-system" and adding args: ["-shared-dev-num","5"] to the container section should work. At least now it is showing 5 available gpu resources.

Can you show a screenshot of where exactly you placed that line in the file? I'm not familiar with this format and not sure where to add it. I'd like both handbrake and plex to have access to the gpu.

 

[ornias]

Can you show a screenshot of where exactly you placed that line in the file? I'm not familiar with this format and not sure where to add it. I'd like both handbrake and plex to have access to the gpu.

If you see the changes in that github issues, place it below the same line inside file opened when executing the command listed by @patmage

 

[Rudi Pittman]

If you see the changes in that github issues, place it below the same line inside file opened when executing the command listed by @patmage

Got it to work. Thank you!