Hi Shatil,

I want to point out that this is not what I am asking for.  I
specifically did not ask for x32 user-land packages.  I am asking for
the kernel option that allows x32 binaries to run.  I believe this is a
very small change.

This change allows for x32 containers and chroot environments.  Having
the kernel configuration for x32 is a necessary step zero to get started
at all.  How to build the x32 containers is a different problem.  One
option, for instance, is using the chroot x32 environment available in
CernVM.

As you can see from the references, we did a number of tests with
high-energy physics applications.  While the difference in running time
is marginal, the difference in memory consumption is always significant:
10%-20%, sometimes even a bit more.  I’m not aware of many other tricks
that provide such big savings in real-world applications.

To summarize, this request is only about the CONFIG_X86_X32 kernel
option.  Nothing else for the time being.  I created a bugzilla record
here: https://bugzilla.redhat.com/show_bug.cgi?id=1310158

Cheers,
Jakob

On 20/02/16 00:58, Shatil Rafiullah wrote:
> We configure our own kernel and I've been down the path of configuring
> x32 there, but afterwards, getting the user land to play nicely is
> another matter.
> 
> Mind you, I was doing this on EL6, and so I needed to introduce a new
> glibc that co-existed harmoniously with the upstream one, and then
> bootstrap and build GCC with x32 support. Because we were evaluating
> software that already existed for x86_64, I patched Yum and RPM to
> accept x32 as an additional architecture.
> 
> After all that, I finally ran evaluations on the software we planned
> to evaluate in x32, and found the performance gains were negligible
> (it was twemcache, which uses tons of pointers, fyi). I highly
> recommend doing your own evaluation first, before requesting such a
> large change to be applied community-wide. It may not be the panacea
> you seek.
>