Autoload SEP Drivers on boot time for compute nodes on HPC Cluster

[Bruno_B]

Hello,

My setup is an HPC Cluster composed of a headnode and several compute nodes.

I am able to build and install the Sampling Drivers on my headnode, and the drivers autoload on boot time.

On my compute nodes, I can build and install the drivers, but it is not persistent on a reboot.

I mount /opt/intel with NFS on my compute nodes. But the nodes boot using a VNFS image built from my headnode image.

I tried reassembling the VNFS image after running ./boot-script successfully, but the nodes booting with this fresh image do not autoload the drivers.

Anyone faced the same situation and has a fix?

Regards,
Bruno.

[yuzhang3_intel]

I am not sure if you can reload VTune driver using the below command line after compute nodes reboot?  If it can, you can add the command line in the script.

$ insmod-sep -r -g vtune

[Bruno_B]

I need to rebuild the drivers in order to do that in the compute node after a reboot. I suspect I'm not provisioning the already built drivers from the headnode.

[yuzhang3_intel]

Are the VTune drivers (*.ko files) not located in the compute node? I suppose when the compute node setup succeeds the first time, all the drivers are located in the current node, right?

Even though you must rebuild the driver after restarting the compute node, can you add the operations about building and loading the driver to one script file to automatically finish the driver installation?

[Bruno_B]

The .ko files are located in the compute node but I have problems provisioning the kernel source files (kernel-devel) in order to build the driver on the client. Trying to solve that now (not in the scope of vtune).

Once I do that I will try the approach of building/loading with a script. Still, that solutions involves a penalty in the time it takes to boot a node.