Client_died on the GPU client

[johnT89]

Ubuntu 10.04

> SMP2 native on i5 750
> GPU2 with wine 1.2 on nvidia 8600gt

WU's on the GPU2 client are usually completed succesfully, but sometimes I find the client stuck on a x percentage and the gpu temperatures at idle levels. If i try to restart the client, it resume from where it left and usually completes the wu without any more issues. Homewer, it is a problem since I leave the pc unattended for whole days.

The 8600gt is not overclocked, just old. Temperatures never go above 75C. Project: 6600 (Run 3, Clone 232, Gen 215)


1) Is it a sign the card is going to die?
2) May it be a wrapper\cuda problem?
3) Is there any risk of returning bad results?

[toTOW]

1) probably no ... but some people reported that p66xx WUs had a lot of problems on low end cards ...
2) maybe ... but I don't know for sure
3) yes ... but when someone returns bad results, someone else will fold the same WU to check the results and to make sure it's not a WU problem.

[Hyperlife]

I see that happen every once in a while with Core 11 WUs on Ubuntu 9.04. Restarting the client fixes the problem for me too. I don't know of any other solution.

So far, I haven't seen that happen with any Core 15 WUs.

[Sidicas]

I have a suspicion of what might be causing this... Something is timing out.. I'm not sure whether it's because the GPU is getting starved of CPU time or if it's the GPU is too busy to complete the work sent to it in time..

I think the best recommendation I can give is for anybody having this problem to decrease their wait times in the wrapper and recompile..They will see higher CPU usage, but I believe that would take care of the timeouts.

It did appear to me that the nvidia drivers themselves are killing the execution of the CUDA calls pushed out to the GPU for being non-responsive.. I'd guess that might just be because it's getting starved for CPU time but it could also be a problem with the Linux nvidia drivers, unfortunately..

It's very difficult for me to replicate this problem reliably on my current hardware.. So it would be great if somebody else could try decreasing their wait times in their wrapper and report back if the problem went away.

Edit: The other thing you can try is have the folding@home GPU core run at the highest priority possible (Nice -19). To be sure it gets the CPU time it needs so it doesn't get it's CUDA code killed for being non-responsive by the nvidia drivers..