Another "Computation Error" thread

You can very quickly rule out problems with the tasks themselves by looking at how many of your wingmen (quorum partners) are also having problems. Nobody else seems to be having a problem with the tasks but I didn't look at all that many. Also it's rather strange that a small percentage of tasks are completing and validating whilst the bulk are failing.

I can't recall ever seeing that particular message before but the mention of "registry keys" seems to suggest some sort of installation issue, either in BOINC itself or the CUDA driver. If it were my machine I would stop BOINC and uninstall it. Your BOINC Data directory will be preserved and can be picked up again when you eventually reinstall. At each major step I'd be rebooting to make sure the system is "refreshed" and also to give it a chance to "complain" if it wants to. Any such complaints might give you a better handle on what the problem actually is.

The next step would be to remove the current CUDA driver. You might need to do some research on whether or not there are any particular tricks to fully removing the old driver. After rebooting, try reinstalling the latest driver from nVidia (you mentioned 266.58 which I think is the latest anyway), once again keeping a close eye on anything unusual during the installation. If all seems normal, reboot for good measure and then reinstall BOINC, making sure your previous installation is detected. This should happen automatically but it doesn't hurt to use the available options to check for yourself. Then enable work fetch when you are ready (I always set NNT (no new tasks) before shutting down so that I have some control over exactly what happens during a restart) and see if the new work you get can be crunched without error with the fresh installation.

If tasks still fail after that, I'd be looking at testing and replacing hardware, probably starting with your GPU, PSU and system RAM to see if you can find something that changes things.

Good luck with your troubleshooting.

Gary, Charles: I have a clue. All of my errors occurred while I was away from the computer. I noticed that one occurred after my daughter used the computer on her user ID, so I got curious. Tested this morning and indeed, the cuda task that was running bombed when I logged off her id and returned to my id. If I suspend GPU processing before going to her id, all is okay when I return to my id. Charles, do you have multiple users set up on your machine? And have you tried Gary's suggestion to reinstall BOINC?

Just found this comment by Mikey in the Getting Started forum:

http://einsteinathome.org/node/195462&nowrap=true#109745
*****(He was taliking about setting up all users on a machine to run BOINC. I haven't tried that, so this item may not apply)
*****
Be aware that if you crunch with a gpu this will NOT work! Yes Boinc itself will work as you described but the gpu will not crunch when switching from user to user. This is a Windows thing and it is what it is, so if you are crunching with a gpu too, all units will error out when you switch from one user to another. In short do not plan to use a gpu for crunching on a machine with multiple users.
*****

All of my other cuda tasks have run to completion since my earlier post because I've suspended GPU processing when I'm away from the machine. I forgot to mention in my first post that I reinstalled the Nvidia drivers when these problems started, using their installer's menu option for what they call a "clean" install (removes all old files first). That didn't help me.

Mikey and Gary, thanks for the help.

Have taught daughter and grandson how to suspend GPU work before they switch users, but they'll forget from time to time. Might have grandson log me off - it's easier for him to remember, but that stops cpu work too. Just posted a wish list item at BOINC site to ask if BOINC Mgr. could detect the user switch and suspend gpu processing for us - not very hopeful that they can or will, but thought I'd try. If they show interest, I'll refer them to your explanation - it helps me. Thought it was interesting that Primegrid has built in a 10 minute delay after an error to help avoid emptying the cache (and they say it's to save load on their server).

I was wrong when I said SETI didn't fail this way. I thought it was okay because in my first week of cuda processing, I ran 72 cuda tasks from them without error. Didn't have one to test when I wrote that note due to their extended outage. Have some tasks now and sacrificed one for a test Indeed, their tasks fail when I switch user. Mikey, I like the idea of more computers in the house - this may be my excuse to build that dream cruncher, and help avoid the pain of a Win 7 upgrade on this one! But it would be cheaper to locate a new battery for her old notebook.

ps sorry about the typos in these posts - that's a combination of trying to type while GPU is crunching and putting off eye surgery too long.