FAQ  •  Login

Additional GPU nodes (re: cudaErrorDevicesUnavailable)

<<

Andrew Seidl

Administrator
Administrator

Posts: 193

Joined: Thu Oct 28, 2010 11:54 am

Unread post Tue Dec 10, 2013 3:30 pm

Additional GPU nodes (re: cudaErrorDevicesUnavailable)

It looks like most of the main GPU nodes are in use by researchers (end of the semester is always busy). We have two more GPU nodes that you can use: euler99 and lagrange. Both have Tesla K20X GPUs and more recent CPUs than the main GPU nodes (euler99: Xeon E5-2630 Sandy Bridge EP , lagrange: Xeon E5-2690v2 Ivy Bridge EP).

To use from Euler, ssh to either euler99 or lagrange. Next, run 'nvidia-smi' and find the device ID# of a GPU that isn't in use (compute mode should say 'default', compute processes at the bottom shouldn't list anything for that ID). Lastly, run 'export CUDA_VISIBLE_DEVICES=#' where # is the device ID for the available GPU. You should now be able to run your code on the selected GPU (which will now show up in your code as device 0).

(All these steps are automatically done on the main GPU machines if you submit via qsub. It's not done for euler99 or lagrange as those two machines are typically only used by one person at a time for benchmarks.)
<<

f13-759-asinha7

Newbie
Newbie

Posts: 45

Joined: Mon Sep 09, 2013 9:12 am

Unread post Wed Dec 11, 2013 4:33 am

Re: Additional GPU nodes (re: cudaErrorDevicesUnavailable)

Hey Andrew,

Do we need to use any particular gencode flag values to compile and run on these machines? I have a previously running code which does not launch the kernel now. I get error 0x8 (cudaErrorInvalidDeviceFunction). Please advise.

Regards
Arindam

PS: I built the code for compute_35 and it tends to work but compute_20 does not.
<<

Andrew Seidl

Administrator
Administrator

Posts: 193

Joined: Thu Oct 28, 2010 11:54 am

Unread post Wed Dec 11, 2013 12:38 pm

Re: Additional GPU nodes (re: cudaErrorDevicesUnavailable)

Correct, it's almost always better to build targeting the compute capability of the devices you'll be running than to use an earlier CC.

NVCC allows you to specify up to three CCs when building (though it will take that much longer to build), with the correct one selected automatically at runtime. I typically build for 2.0, 3.0, and 3.5.
<<

f13-759-cheemalamarr

Newbie
Newbie

Posts: 19

Joined: Tue Sep 10, 2013 11:01 pm

Unread post Fri Dec 13, 2013 1:25 pm

Re: Additional GPU nodes (re: cudaErrorDevicesUnavailable)

Thank you Dan and Andrew for getting back to us on this issue.


Regards,
Sai

Return to ME759 Fall 2013: High Performance Computing

Who is online

Users browsing this forum: No registered users and 1 guest

Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group.
Designed by ST Software.