some of us do but that has be deduced to problems with the driver itself and posted on nvidia...

were talking about app errors wit the app and boinc...

and some these are related to the fact that gpugrid is still built upon 2.1 cuda which is in the upgrade department right now.

Thanks for that info..would you then be interested in the fact that rolling back to 182.08 then 185.85,186's and 190's they all display the same behaviour
Forcing Riva and locking 3D has no impact nor does 6.6.36 to 6.6.37.(this on vista 64 and 2* gtx260 (192's) Cuda 2.1.

Perhaps the most annoying thing for me though is that errors can occur from just a few seconds out to almost 100% complete.

I think I may just try with xp again.

Mike

It worked on mine until it downloaded a new task and it errored out on the download " Cuda.dll Unexpected size " or something like that.

Hello Admin,
Would it be possible to do as Seti and have it to where the 2 DLL files for Cuda 2.3 could be " dropped " into the GPU folder and replace the 2.2 files?
This is possible on the Seti optimized apps. and works great.
For the people that 190.. drivers are running problem free this could give a small added boost.
I tried it one time and it would not work because the "file size " was incorrect and it errored out.
If the software did not look at the file size it would maybe work.
It may also help the developers get a jump on any issues with 2.3 also.

Just a question.
Thanks

That sounded like an interesting question, so I'm giving it a try on two hosts.

Preliminary observation - makes no difference at all, but both tasks are still below 10% - it'll be tomorrow morning (UTC) before I get a definitive answer. I'm running this on my two Q6600/9800GT hosts (XP SP3), if anyone wants to watch.

I can only presume that GPUGrid relies much less on Fourier Transforms than SETI - it was fft that made the difference at SETI.

http://boincfaq.mundayweb.com/index.php?language=1&view=91

Please check if you can convince BOINC with this option to ignore the incorrect size and proceed with 2.3 libs:

<dont_check_file_sizes>0|1</dont_check_file_sizes>

First task with 2.3 runtime DLLs (1041263) errored out after 14 hours with "Cuda error: Kernel [frc_sum_kernel_dihed] failed in file 'force.cu' in line 313 : unspecified launch failure."

Task on second machine, and replacement for this one, still running: but proceed with caution.

It worked on mine until it downloaded a new task and it errored out on the download " Cuda.dll Unexpected size " or something like that.

Hello Admin,
Would it be possible to do as Seti and have it to where the 2 DLL files for Cuda 2.3 could be " dropped " into the GPU folder and replace the 2.2 files?
This is possible on the Seti optimized apps. and works great.
For the people that 190.. drivers are running problem free this could give a small added boost.
I tried it one time and it would not work because the "file size " was incorrect and it errored out.
If the software did not look at the file size it would maybe work.
It may also help the developers get a jump on any issues with 2.3 also.

Just a question.
Thanks

That sounded like an interesting question, so I'm giving it a try on two hosts.

Preliminary observation - makes no difference at all, but both tasks are still below 10% - it'll be tomorrow morning (UTC) before I get a definitive answer. I'm running this on my two Q6600/9800GT hosts (XP SP3), if anyone wants to watch.

I can only presume that GPUGrid relies much less on Fourier Transforms than SETI - it was fft that made the difference at SETI.

You'd need to use an app_info file. That way you'd bypass BOINC checking the signatures on the files.

---

BOINC blog

Would just adding this line into the CC_Config file work?

<dont_check_file_sizes>0|1</dont_check_file_sizes>

I can only presume that GPUGrid relies much less on Fourier Transforms than SETI - it was fft that made the difference at SETI.

You'd need to use an app_info file. That way you'd bypass BOINC checking the signatures on the files.

Strangely, the CUDA RT DLL is signed, but the FFT isn't. Seriously, does anyone know whether the mathematics here requires Fourier Transforms? If not, it's not worth chasing this.

And yes, I am using an app_info.xml

1040687 completed and validated using the 2.3 DLLs. Took slightly longer than previous tasks on that box, but got slightly more credit - must have been an tougher job. Hourly rate exactly the same. Conclusion? It's not worth the hassle. Reverting to stock so I can get the new compiles as they come out.

We had already optimized the FFT our selves, so no much speedup will be seen using 2.3.
By the way, we have updated the application to cuda 2.2.

gdf

We had already optimized the FFT our selves, so no much speedup will be seen using 2.3.
By the way, we have updated the application to cuda 2.2.

gdf

thanks for updating, although last night i grabbed 40 seti cuda wu b/c i knew this change would happen today, and i just wanted to see what feedback was coming out from here... on the 6.67 app.

We had already optimized the FFT our selves, so no much speedup will be seen using 2.3.

Ah well, experimental proof is always good to go alongside the theoretical conjecture.... :-)

By the way, we have updated the application to cuda 2.2.

gdf

And started distributing the 2.2 DLLs. Nice workround for the 'same name, different version' problem that BOINC handles so badly.

Sorry, I tried to unravel my app_info.xml without doing a full project reset - and failed. I've released the WU back into the wild for someone else to do.

And ended up with a reset project anyway. DCF 1.000000, <rsc_fpops_est>250000000000000 works out to 01:08:53 on this host, where it should be about 16 hours.

I could edit client_stats again, or you could - pretty please - put a more realistic estimate into the workunit generator?

We will start to increase by a factor 2. Then we need to wait a little.

gdf

I tried the Grid Wu's again last night when I seen the Application went to CUDA 2.2 but I'm still getting error's on at least 2 Box's with the 190.38 Drivers & 6.6.36 Client, same 2 Box's that have given me the most trouble since Updating to the 190.38 Drivers that didn't have any trouble before Upgrading.

I tried the Grid Wu's again last night when I seen the Application went to CUDA 2.2 but I'm still getting error's on at least 2 Box's with the 190.38 Drivers & 6.6.36 Client, same 2 Box's that have given me the most trouble since Updating to the 190.38 Drivers that didn't have any trouble before Upgrading.

So which two boxes are they, and what graphics cards do they have installed?

Consensus seems to be settling down on 'G92-based cards run the new drivers OK, g200-based cards have problems'.

I tried the Grid Wu's again last night when I seen the Application went to CUDA 2.2 but I'm still getting error's on at least 2 Box's with the 190.38 Drivers & 6.6.36 Client, same 2 Box's that have given me the most trouble since Updating to the 190.38 Drivers that didn't have any trouble before Upgrading.

So which two boxes are they, and what graphics cards do they have installed?

Consensus seems to be settling down on 'G92-based cards run the new drivers OK, g200-based cards have problems'.

Right now it's Host ID 17909 & Host ID 38967 ... On the 17909 Host both 260 Cores were running @ 300/Core & 400/Memory when I checked it, On Host 38967 1 of the 260 Cores was running @ 300/Core & 400/Memory while the other Core was running 650/Core & 1100/Memory.

Both Box's had been running the Collatz Projects Wu's for the last 3-4 Days with no Error's or Re-setting of the Cores to Lower Speeds. I did a complete Uninstall of the NVIDIA Drivers again on both Box's, then Re-installed them to each 260 Card thru the Device Manager on both Box's, did the Riva Tuner Fix thingy on both Box's gave them both a Big Hug & Kiss to let them know they were still Loved ... hahaha ... and sent them on their way running the Grid Wu's again to see if that fixes them.

If they still spit out the errors then it's back to Collatz with them, so far the rest of my Pharm has behaved it's self and hasn't had any errors since last night.

What is the difference in the 'G92-based cards & the g200-based card ??? I just thought they all were 200 Series Cards. Funny thing is I have 4 Box's with 2 BFG GTX 260's each. 2 of the Box's seem to run the Wu's Okay but then the other 2 BFG GTX 260 Box's just spit out errors. All 8 Cards are the BFG 216 Shader OCX Version Type Cards ... Weird

NOTE: One thing I have noticed is that if one of the Cores drops or resets it's self to a lower running speed but the other one doesn't your still just going to get error's off both Cores. So the Speed drop must have an effect on both Cores for some reason or other.

I did a RMA on a BFG GTX 260 Card the same as the ones I have now just last week for the same reason of the Speed Drop. Since I got it's replacement back it's run okay, could just be something with the BFG Cards but I don't think so with all the other people reporting problems too.

it does not seem that you have successfully downgraded to 185.xx:
http://www.gpugrid.net/show_host_detail.php?hostid=17909

Your are having problems with the CUDA FFT library, which also seem to suggest that you are using the wrong driver. Please check.

Do you have the toolkit installed, if yes uninstall it. I am not sure if Windows would choose system installed libs rather than the one we provide.

gdf

it does not seem that you have successfully downgraded to 185.xx:
http://www.gpugrid.net/show_host_detail.php?hostid=17909

Your are having problems with the CUDA FFT library, which also seem to suggest that you are using the wrong driver. Please check.

Do you have the toolkit installed, if yes uninstall it. I am not sure if Windows would choose system installed libs rather than the one we provide.

gdf

Hmmmmmm, but I tried that already (Downgrade to 185.18 I believe) & it din't work, I still got the errors, that and all my other Box's are running the 190.38 Drivers. I'll wait and see if they run okay now with the uninstall & reinstall I did with them an hour or so ago. If they still get the errors I'll try the 185.xx Drivers again.

No, I don't have the Toolkit installed on any of my Box's, did at first when I joined the Project but then un-installed them when I realized I didn't need them.

GDF, do you mean Downgrade to 185.xx Drivers or 186.xx Drivers, I think I tried the 186.18 Drivers but they didn't stop the Errors or the Cards from Dropping speed after just a few minutes of running sometimes, other times it took an hour or so before they dropped their speed.

PS: Crap, looks like something is wrong with one of the Box's already, 1 of the Wu's isn't Progressing according to BOINCView, will have to go out to the Garage & see what's up with it. Think I'll just go ahead and put the 185.xx Drivers in it & try them.

Looks like it did a Benchmark Test & upon Resuming the Wu's one of them erred out ... Host ID 17909 One Core & 300/Core & 400/Memory & One Core @ 650/Core & 1100/Memory ... Going to the 185.85 Drivers & see if they work or not ...

8/1/2009 8:14:24 AM Running CPU benchmarks
009 8:14:24 AM Suspending computation - running CPU benchmarks
8/1/2009 8:14:55 AM Benchmark results:
8/1/2009 8:14:55 AM Number of CPUs: 4
8/1/2009 8:14:55 AM 2275 floating point MIPS (Whetstone) per CPU
8/1/2009 8:14:55 AM 6944 integer MIPS (Dhrystone) per CPU
8/1/2009 8:14:56 AM Resuming computation
8/1/2009 8:29:00 AM AQUA@home Sending scheduler request: To fetch work.
8/1/2009 8:29:00 AM AQUA@home Requesting new tasks for CPU
8/1/2009 8:29:05 AM AQUA@home Scheduler request completed: got 0 new tasks
8/1/2009 8:29:05 AM AQUA@home Message from server: No work sent
8/1/2009 8:29:05 AM AQUA@home Message from server: (reached limit of 2 CPU tasks)
8/1/2009 9:10:05 AM GPUGRID Computation for task 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0 finished
8/1/2009 9:10:05 AM GPUGRID Output file 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0_1 for task 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0 absent
8/1/2009 9:10:05 AM GPUGRID Output file 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0_2 for task 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0 absent
8/1/2009 9:10:05 AM GPUGRID Output file 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0_3 for task 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0 absent
8/1/2009 9:10:06 AM Collatz Conjecture Restarting task collatz_1248745476_67361_1 using collatz version 110
8/1/2009 9:10:08 AM GPUGRID Started upload of 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0_0
8/1/2009 9:10:08 AM GPUGRID Started upload of 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0_4
8/1/2009 9:10:12 AM GPUGRID Finished upload of 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0_0
8/1/2009 9:11:37 AM GPUGRID Finished upload of 93-KASHIF_HIVPR_twomons_ba2-14-100-RND4610_0_4

PS: Have the 185'85 Drivers installed now, did find a couple of Odd Ball nv.dll's in the system 32 Directory, or at least i thought they were Odd Ball ones because the referred to an older Driver Version so I canned them before installing the 185.85 Drivers.

I also Aborted the running Grid Wu because I haven't had a Wu finish yet that was already running when one erred out, the remaining Wu would always error out eventually too so I figured 2 Fresh ones would be better. Am running both of them now with the 185.85 Drivers & 6.6.36 Client.