Take the 2-minute tour ×
Server Fault is a question and answer site for professional system and network administrators. It's 100% free, no registration required.

We use a lot of GPGPU computing (mostly with CUDA, but some OpenCL). Often when users are running code, the code errors out with a memory error on only one of our hosts. I suspect one of the cards is faulty. Sometimes it brings down the whole system and sometimes the program just bombs out.

What are the easiest, fastest, and most thorough ways to fully test GPUs for possible failures?

I know there are programs that are part of nvidia's CUDA SDK:

   deviceQuery
   nvidia-smi

But I need something much more thorough. Suggestions? Experiences?

share|improve this question

2 Answers 2

up vote 1 down vote accepted

The de facto standard seems to be CUDA GPU Memtest. As @c2h5oh mentioned, it looks like it's based on memtest86 test patterns so I'm sure it does a good job. It runs relatively quickly on the high end GPUs I'm testing (30 minutes on a Quadro 6000 and 20 minutes on a Tesla C2075). It runs inside the OS (unlike, memtest) so monitoring is a bit different. You'll probably want to output stdout and stderr to a file to look at later. So consider running it something like this in case you lose your terminal output you can look up what the tests found:

cuda_memtest 2>cuda_memtest.stderr 1>cuda_memtest.stdout &
tail -f cuda_memtest.stdout &
tail -f cuda_memtest.stderr &

You'll also want to make sure that no one is using the system and/or cards. You can set the GPUs to exclusive mode using:

nvidia-smi --compute-mode=EXCLUSIVE_PROCESS

Here are some of the output from sample runs of both the Quadro and the Tesla in case you're interested in what test info is given:

[09/07/2012 11:56:22][hydro][0]:Running cuda memtest, version 1.2.2
[09/07/2012 11:56:23][hydro][0]:Warning: Getting serial number failed
[09/07/2012 11:56:23][hydro][0]:NVRM version: NVIDIA UNIX x86_64 Kernel Module  295.41  Fri Apr  6 23:18:58 PDT 2012
[09/07/2012 11:56:23][hydro][0]:num_gpus=1
[09/07/2012 11:56:23][hydro][0]:Device name=Quadro 6000, global memory size=6441992192
[09/07/2012 11:56:23][hydro][0]:major=2, minor=0
[09/07/2012 11:56:24][hydro][0]:Attached to device 0 successfully.
[09/07/2012 11:56:24][hydro][0]:Allocated 6040 MB
[09/07/2012 11:56:24][hydro][0]:Test0 [Walking 1 bit]
[09/07/2012 11:56:30][hydro][0]:Test0 finished in 5.7 seconds
[09/07/2012 11:56:30][hydro][0]:Test1 [Own address test]
[09/07/2012 11:56:33][hydro][0]:Test1 finished in 3.5 seconds
[09/07/2012 11:56:33][hydro][0]:Test2 [Moving inversions, ones&zeros]
[09/07/2012 11:57:05][hydro][0]:Test2 finished in 32.3 seconds
[09/07/2012 11:57:05][hydro][0]:Test3 [Moving inversions, 8 bit pat]
[09/07/2012 11:57:37][hydro][0]:Test3 finished in 31.9 seconds
[09/07/2012 11:57:37][hydro][0]:Test4 [Moving inversions, random pattern]
[09/07/2012 11:57:53][hydro][0]:Test4 finished in 15.9 seconds
[09/07/2012 11:57:53][hydro][0]:Test5 [Block move, 64 moves]
[09/07/2012 11:57:59][hydro][0]:Test5 finished in 6.3 seconds
[09/07/2012 11:57:59][hydro][0]:Test6 [Moving inversions, 32 bit pat]
[09/07/2012 12:18:46][hydro][0]:Test6 finished in 1246.6 seconds
[09/07/2012 12:18:46][hydro][0]:Test7 [Random number sequence]
[09/07/2012 12:19:06][hydro][0]:Test7 finished in 19.8 seconds
[09/07/2012 12:19:06][hydro][0]:Test8 [Modulo 20, random pattern]
[09/07/2012 12:19:06][hydro][0]:test8[mod test]: p1=0x13472f5f, p2=0xecb8d0a0
[09/07/2012 12:20:34][hydro][0]:Test8 finished in 88.0 seconds
[09/07/2012 12:20:34][hydro][0]:Test10 [Memory stress test]
[09/07/2012 12:20:34][hydro][0]:Test10 with pattern=0x55f6c69858704128
[09/07/2012 12:21:11][hydro][0]:Test10 finished in 36.8 seconds
[09/07/2012 12:21:11][hydro][0]:Test0 [Walking 1 bit]
[09/07/2012 12:21:16][hydro][0]:Test0 finished in 5.8 seconds



[09/06/2012 18:49:07][hydro][0]:Running cuda memtest, version 1.2.2
[09/06/2012 18:49:10][hydro][0]:Warning: Getting serial number failed
[09/06/2012 18:49:10][hydro][0]:NVRM version: NVIDIA UNIX x86_64 Kernel Module  295.41  Fri Apr  6 23:18:58 PDT 2012
[09/06/2012 18:49:10][hydro][0]:num_gpus=1
[09/06/2012 18:49:10][hydro][0]:Device name=Tesla C2075, global memory size=5636292608
[09/06/2012 18:49:10][hydro][0]:major=2, minor=0
[09/06/2012 18:49:11][hydro][0]:Attached to device 0 successfully.
[09/06/2012 18:49:11][hydro][0]:Allocated 5273 MB
[09/06/2012 18:49:11][hydro][0]:Test0 [Walking 1 bit]
[09/06/2012 18:49:22][hydro][0]:Test0 finished in 11.1 seconds
[09/06/2012 18:49:22][hydro][0]:Test1 [Own address test]
[09/06/2012 18:49:25][hydro][0]:Test1 finished in 3.1 seconds
[09/06/2012 18:49:25][hydro][0]:Test2 [Moving inversions, ones&zeros]
[09/06/2012 18:49:52][hydro][0]:Test2 finished in 27.4 seconds
[09/06/2012 18:49:52][hydro][0]:Test3 [Moving inversions, 8 bit pat]
[09/06/2012 18:50:20][hydro][0]:Test3 finished in 27.9 seconds
[09/06/2012 18:50:20][hydro][0]:Test4 [Moving inversions, random pattern]
[09/06/2012 18:50:34][hydro][0]:Test4 finished in 13.7 seconds
[09/06/2012 18:50:34][hydro][0]:Test5 [Block move, 64 moves]
[09/06/2012 18:50:39][hydro][0]:Test5 finished in 5.5 seconds
[09/06/2012 18:50:39][hydro][0]:Test6 [Moving inversions, 32 bit pat]
[09/06/2012 19:08:34][hydro][0]:Test6 finished in 1074.9 seconds
[09/06/2012 19:08:34][hydro][0]:Test7 [Random number sequence]
[09/06/2012 19:08:51][hydro][0]:Test7 finished in 17.1 seconds
[09/06/2012 19:08:51][hydro][0]:Test8 [Modulo 20, random pattern]
[09/06/2012 19:08:51][hydro][0]:test8[mod test]: p1=0x63136646, p2=0x9cec99b9
[09/06/2012 19:10:10][hydro][0]:Test8 finished in 78.4 seconds
[09/06/2012 19:10:10][hydro][0]:Test10 [Memory stress test]
[09/06/2012 19:10:10][hydro][0]:Test10 with pattern=0x26341d134a89ac2b
[09/06/2012 19:10:39][hydro][0]:Test10 finished in 29.0 seconds
share|improve this answer

Google: Memtest + GPU: any of the 3 first results seem to be a valid answer. No personal experience.

http://sourceforge.net/projects/cudagpumemtest/

http://www.softpedia.com/get/Tweak/Memory-Tweak/CUDA-MemTest.shtml

https://simtk.org/home/memtest/

share|improve this answer
    
Of course I googled it. I was wondering if anyone knew how all these compared (easiest, fastest, most thorough) for finding errors, testing the features of different cards, etc. –  Andrew Case Jul 3 '12 at 22:31
1  
The first one claims to be replicating memtest86/memtest86+ testing patterns on GPU and that's recommendation enough - those tests will be thorough (and if the tool is anything like memtest86/+ there will be almost no user interaction - start & watch progress). Since you specifically requested memory testing none of the above will test anything but memory. –  c2h5oh Jul 4 '12 at 0:45

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.