I want to create a boot disk image with all the necessary software for my environment installed, beyond what is provided in one of the deep learning images. It seems like installing the drivers on one instance and persisting that boot disk image, gives you a boot disk that can be used for other instance types (following this - the "How can I reduce the startup time?" section).
Are there any concerns from doing this? Wouldn't the drivers and other software be installed in a way that might be particular to the GPU (or other hardware) of that first instance? Would this either create issues or suboptimal performance? If the latter, would it be a meaningful difference versus having a startup script install the software if I'm going to use a different instance type?
Are there any concerns from doing this?
Probably only one, you will have, time to time, do it again in order to update to the latest version of DLVM. BTW on the latest versions of DLVM we are no longer rebooting the VM after installing Nvidia driver. So the first boot is now faster, but installing part is still there.
Wouldn't the drivers and other software be installed in a way that might be particular to the GPU (or other hardware) of that first instance? Would this either create issues or suboptimal performance? If the latter, would it be a meaningful difference versus having a startup script install the software if I'm going to use a different instance type?
We are using same driver for all GPUs (on the latest DLVMs it is 410.*), so there is no difference here. Reason why drivers installed dynamically is purely due to license that prevents drivers to be pre-baked.