Seg Faults with GTX 295 in compute exclusive mode

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi Everyone,

I currently have a GTX 295 card in my debian Squeeze x86_64 system that shows up as 2 GPUs using CUDA 2.2 with the 185.18.36 driver. I set each GPU to compute-exclusive mode, so that a free GPU would be assigned to each independent kernel I launch. I tried testing this configuration by running two independent instances of the same program I wrote, which works using 1 GPU in default mode. When I run the first instance of my program, and then run a second instance in another terminal shortly after, the first instance of my program segmentation faults and I get the following in dmesg:

[14091.451808] test_cuda[17877]: segfault at 8 ip 00007fe14792a3a7 sp 00007fffb83dab80 error 4 in libcudart.so.2.2[7fe14790d000+3e000]
[14094.485504] NVRM: Xid (0005:00): 13, 0001 00000000 000050c0 00000368 00000000 00000080
[14097.028211] NVRM: Xid (0006:00): 13, 0001 00000000 000050c0 00000368 00000000 00000080

Would anyone know why I am getting this error? Please let me know if you have any ideas!

Jonathan

2

That looks like a driver problem to me. Those Xid 13 messages suggest that the driver has been forced to do a hard reset on both devices. You might want post a complete repro case and file a proper bug report. You might also consider upgrading to the 190 series drivers and CUDA 2.3 and see whether the problem goes away.

3

This is all good advice. Also, are you checking all of your error codes? Can you post a repro?

4

Hey guys, thanks for the quick response!

I was able to get it working using the 190.18 beta drivers that are recommended for use with CUDA 2.3. However, when I try to use the newer 190.32 and 190.36 drivers, I continue to get seg faults in my code.

Is this something I should submit a bug report for? If so, what is the process to do so?

Jonathan

5

You post a repro case so I can try it. :) (alternately, you PM me source so I can try it)