[ale] CentOS 7 and NVIDIA

[Michael B. Trausch]

On 07/25/2014 05:28 PM, Jim Kinney wrote:
> Don't mix 'em yet.
>
> for the efficient (i.e. busy) admin, RPMFusion is not yet online with v7.
> So no NVIDIA modules to run a k4000 or a Tesla GPU.
>
> The official Nvidia installer pukes during the build with a failed kernel
> test in autoconf and suggests a make oldconfig && make prepare
>
> which also fails in the 3.10.0-123.4.4 kernel tree with a failed call to
> arch/x86/syscall/syscall_32.tbl
>
> There is_no_  32 bit support at all in RHEL7. So what's with a 32-bit sys
> lookup table?
>
> Running kernel works fine but building new stuff fails instantly. ugh.
You can install 32-bit headers to make the build happy, and GCC can 
build the 32-bit module. That said, you'd have to not use nVidia's 
script to do it, but do it by hand. You can probably patch it to just 
avoid it as well—most problems I've had building nvidia software on 
"unsupported" environments often amounted to patching 10 or less lines 
to get it to correctly and successfully build.

Having now learned The Way of Red Hat, I also spend much less time 
building and deploying binaries than I used to. Build once, write spec 
file, build again with split packages, upload to repo, and just yum 
install everywhere.

And, as long as I manage 100% of the mirrors and repositories, yum 
never, ever breaks! It's all about finding the tricks!

(Sorry, though, I can't help with nvidia... I don't think I have any 
more nvidia hardware left. I won't use a graphics display unless it's 
got KMS-enabled drivers in-tree that perform well... and I don't do 
anything that requires massive vector processing since most of my work 
is programming or data manipulation.)

You might try doing the build on a F19 or F20 system and simply copying 
the 64-bit modules to the R7/C7 system. That might actually work, since 
they're similar enough.

— Mike