cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

thehighhat
Journeyman III

headless opencl / stream

opencl running on gpu without display attached (linux)

with the 2.01 release and the catalyst 10.1 drivers, is it yet possible to run openCL codes on the GPU, without a display?

 

linux x86_64

fglrx from 10.1 driver package

stream sdk 2.01

no display attached to pcie card

run apps remotely on the GPU

 

even better:

init pcie card even if not primary devices (ie multiple cards on one motherboard)

none have active X display or console

run openCL apps on chosen GPUs

installation paths for driver and libs that allow for nvidia & ati to coexist on same machine

io-srv access to pcie via kvm (woot!) for accelerated GPU... <-- that's a money maker!

 

 

but still good enough...

initialize GPU to only *think* it is connected to display

0 Likes
8 Replies
nou
Exemplar

i can run OpenCL application via ssh. without attached monitor. but you must have running Xserver with logged session. then i can log in via ssh export DISPLAY=:0 and run OpenCL application. you must have running Xserver

http://developer.amd.com/support/KnowledgeBase/Lists/KnowledgeBase/DispForm.aspx?ID=19

0 Likes

Originally posted by: nou i can run OpenCL application via ssh. without attached monitor. but you must have running Xserver with logged session. then i can log in via ssh export DISPLAY=:0 and run OpenCL application. you must have running Xserver

http://developer.amd.com/support/KnowledgeBase/Lists/KnowledgeBase/DispForm.aspx?ID=19

 

Yes, I remember reading about this approach.

 

My goal is to avoid initializing X like this at all.  No live Xserver running.  No export hacks.

 

Two problems surface...  currently the ATI  drivers require Xserver to init the card.  This prevents multiple cards from different vendors from working together except in very peculiar configurations.  Second, this limits utiltiy for remote installations... cannot run massively parallel on cluster of mobos with pcie cards. 

I can also imagine that the user that inits xserver will be only one to get access to card...  if i build a rack of machines (100s of tflops) and need to mpi push computation out to gpus, it is easy on cuda on nvidia... totally impractical with amd.

0 Likes

Xserver, puke!  Highhat, thanks for putting this up.  My definition of Headless would allow any program, started by a login or at boot, i.e. HTTP server, to directly interface.  This work around is not production worthy, IMHO.

0 Likes

Originally posted by: jcpalmer Xserver, puke!  Highhat, thanks for putting this up.  My definition of Headless would allow any program, started by a login or at boot, i.e. HTTP server, to directly interface.  This work around is not production worthy, IMHO.

 

I agree. There must be a better way. Isn't there a supercomputer where a cluster of 5000 Radeon GPU nodes are utilized? Do they run 5000 Xservers in the cluster as well??

(EDIT: 5000 as in the counting number of GPUs, although it's probably the series number )

0 Likes

i think it run only 2500 Xservers as one node have 2 GPU.

0 Likes

Originally posted by: nou i think it run only 2500 Xservers as one node have 2 GPU.

 

No kidding! They're not doing it a better way?

0 Likes

Did you find a way to run opencl applications without logging into X, only via ssh?

0 Likes

I just got a GTX 480 and found that the nvidia driver module needs to believe it has an attached display and screen to unlock. CUDA and OpenCL won't work until after starting an X server (which may then subsequently exit).

See: http://devforums.amd.com/devforum/messageview.cfm?catid=328&threadid=117913&enterthread=y

0 Likes