X / CUDA on CentOS, multi-gpu servers

[Ben Scott]

On Mon, Oct 14, 2013 at 3:14 PM, Bruce Labitt <Bruce.Labitt at autoliv.com> wrote:
> ... issue getting X running on our multi-gpu server with CUDA. ...
> What would it take to get X and CUDA to play nice?  How much time?  Anyone available to do this?
> We want to remote in to the server and have the display directed to our local machines.

  General terminology: X display server = Thing that provides a
graphics display device.  X client = program which uses a graphics
display device.  Programs like Firefox and CAD are X clients.  On
Linux these days, the X display server is most commonly provided by
the X.org group, possibly aided by some software/drivers provided by
the video card maker.  Do not fall into the trap of thinking an X
display server musty run on a big computer in a closet everyone calls
"the server".  Do not fall into the trap of thinking an X client must
run on the small computer on your desk.

  For your scenario:

  For clarity of discussion, let's suppose your server's name is FRED.

  For the use case you describe, you do not need to run an X display
server on the FRED at all.  The X client(s) will run on FRED, and the
X display server will run on your workstation/terminal.  So for that
use case, do not bother configuring or running an X display server on
FRED; it's just a waste of resources.

  Be aware that GPU-accelerated graphics are generally only available
on a local display.  This is true for all the common protocols (X,
RDP, VNC).  GPU-accelerated non-display computation can still be done
remotely, though.  This would be appropriate if a lot of heavy math
(e.g., a complex simulation) needs to be done before anything can even
be displayed in the first place.

-- Ben