nV News Forums

 
 

nV News Forums (http://www.nvnews.net/vbulletin/index.php)
-   NVIDIA Linux (http://www.nvnews.net/vbulletin/forumdisplay.php?f=14)
-   -   Quadro 600 SLES11SP1 variety of problems, driver suspect (http://www.nvnews.net/vbulletin/showthread.php?t=178416)

rburgess 04-16-12 02:05 PM

Quadro 600 SLES11SP1 variety of problems, driver suspect
 
2 Attachment(s)
Hello,

I have a Q600 that will run X, but cudasdk examples run very inconsistently or not at all (nbody, bandwidthTest, etc.), driver loads slowly (15-30 seconds) and inconsistently.

290.10 driver almost never loads, gives error:
NVIDIA: could not open the device file /dev/nvidia0 (Input/output error).
Failed to initialize NVML: Unknown Error

295.33 intermittently loads, gives same error as above. Also noticed that bandwidthTest reports:
cudaGetDeviceProperties returned 10
-> invalid device ordinal

295.40 seems to fix the driver loading problem, but cudasdk samples do not always run; they often get stuck in a loop (30-80% of the time). I do not see anything different when comparing an strace -f from a good run and a bad run; the bad one just gets stuck looping this:
futex(0x64fd20, FUTEX_WAIT_PRIVATE, 2, NULL) = -1 EAGAIN (Resource temporarily unavailable)
futex(0x64fd20, FUTEX_WAKE_PRIVATE, 1) = 0

An FX1800 in the same system does not have the same problems. I have recompiled the samples on this system with no change. I see this card should have been supported for quite some time. I've also tried older drivers (275, 280) with no luck.

Does anyone have any suggestions for further troubleshooting?

rburgess 04-17-12 09:45 AM

Re: Quadro 600 SLES11SP1 variety of problems, driver suspect
 
275.21 drivers load in only 2s, but they appear to be too old to run CUDA sdk examples:
cudaGetDeviceProperties returned 35
-> CUDA driver version is insufficient for CUDA runtime version
[bandwidthTest] test results...
FAILED

rburgess 04-18-12 08:02 AM

Re: Quadro 600 SLES11SP1 variety of problems, driver suspect
 
It looks like there may be a 64GB system memory limit for this card with current drivers. I can pin the tests to any CPU in the system, as long as I only use memory from CPU0 (64GB of 256GB). Pulling memory from each CPU so the system total is 64GB (16GB/ea) allows tests to run in memory attached to all CPUs. Confirmed adding memory to CPU0 (32GB; system total 80GB) makes tests against CPU3 memory fail.

Is there a better contact than linux-bugs@nvidia.com for reporting these problems? I don't even get an automated response from that address.

sandipt 04-19-12 07:35 AM

Re: Quadro 600 SLES11SP1 variety of problems, driver suspect
 
what versions of cuda sdk, cuda toolkit and gpu driver you are using?

rburgess 04-19-12 11:48 AM

Re: Quadro 600 SLES11SP1 variety of problems, driver suspect
 
Quote:

Originally Posted by sandipt (Post 2547444)
what versions of cuda sdk, cuda toolkit and gpu driver you are using?

sdk/toolkit 4.1.21 and 4.1.28 have been tried; 295.40 driver right now, but others have been tried, as listed above.


All times are GMT -5. The time now is 09:00 AM.

Powered by vBulletin® Version 3.7.1
Copyright ©2000 - 2014, Jelsoft Enterprises Ltd.
Copyright 1998 - 2014, nV News.