CUDA App - Not Threadsafe?
log in

Advanced search

Message boards : Windows : CUDA App - Not Threadsafe?

Author Message
EPAstor
Send message
Joined: 24 Jan 10
Posts: 4
Credit: 19,194,722
RAC: 0
Message 13993 - Posted: 14 May 2012, 0:03:02 UTC

I don't know about anyone else, but I've been participating in this project for some time, using my nVidia GPU to contribute more results. Ever since the nVidia 270 series drivers, I've been having random crashes after I leave my system unused for a short time, then move the mouse. The result is usually a graphics driver crash, sometimes bad enough to take down my entire Windows 7 system.

It turns out that this appears to be related to BOINC - and that this project COULD make a workaround to fix the issue! However, it hasn't. Any chance this problem can get fixed soon? I'm a programmer myself, and happy to lend a hand - but I suspect the change is either relatively minor, or that the people over at the lunatics forum (who've studied this problem - my source was actually http://lunatics.kwsn.net/1-discussion-forum/recent-driver-cuda-safe-project-list.0.html) could help describe the exact workaround needed to allow BOINC to safely snooze the process once the computer is back in use.

Profile mikey
Avatar
Send message
Joined: 11 Aug 09
Posts: 3245
Credit: 1,698,651,825
RAC: 5,024,826
Message 13997 - Posted: 14 May 2012, 10:45:06 UTC - in response to Message 13993.

I don't know about anyone else, but I've been participating in this project for some time, using my nVidia GPU to contribute more results. Ever since the nVidia 270 series drivers, I've been having random crashes after I leave my system unused for a short time, then move the mouse. The result is usually a graphics driver crash, sometimes bad enough to take down my entire Windows 7 system.

It turns out that this appears to be related to BOINC - and that this project COULD make a workaround to fix the issue! However, it hasn't. Any chance this problem can get fixed soon? I'm a programmer myself, and happy to lend a hand - but I suspect the change is either relatively minor, or that the people over at the lunatics forum (who've studied this problem - my source was actually http://lunatics.kwsn.net/1-discussion-forum/recent-driver-cuda-safe-project-list.0.html) could help describe the exact workaround needed to allow BOINC to safely snooze the process once the computer is back in use.


To crunch faster use the power button on your monitor NOT the screen saver blanking. Also both the 295 and 296 series of the Nvidia drivers have problems with the screen saver blanking and Boinc, the later versions do not. Remember the Nvidia drivers are mainly for gamers, we crunchers just kind of tag along for the ride! Also once you find a driver version that works and is steady, don't upgrade until you need to, unless of course you game too!

Claggy
Send message
Joined: 27 Sep 09
Posts: 288
Credit: 14,320,498
RAC: 0
Message 14001 - Posted: 14 May 2012, 19:29:17 UTC - in response to Message 13993.
Last modified: 14 May 2012, 19:46:41 UTC

When i was running the Collatz Cuda app on my 128Mb 8400M GS, when i suspended GPU computation, it would always take 10 seconds for the app to exit (with Boinc 6.10.58),

Raistmer's OpenCL Nvidia Astropulse app was sometimes causing driver restarts when suspending (with +270.xx drivers) after about 15 seconds, and wasn't putting anything in the stderr.txt about exiting (it already had threadsafe api at this time),
i remembered about the Collatz Cuda app's behavour and worked out that Boinc was killing the app at 10 seconds (or 15 seconds with Boinc 6.12.x and later) because it wasn't exiting when told to do so,
I managed to get Raistmer to do the exit check more often (I think he moved it from the outer loop to the middle loop), now app exits in well under a second, and doesn't cause a driver restart, the same could be done to the Collatz Cuda apps,

Claggy

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 14005 - Posted: 15 May 2012, 14:38:22 UTC - in response to Message 14001.

When i was running the Collatz Cuda app on my 128Mb 8400M GS, when i suspended GPU computation, it would always take 10 seconds for the app to exit (with Boinc 6.10.58),

Raistmer's OpenCL Nvidia Astropulse app was sometimes causing driver restarts when suspending (with +270.xx drivers) after about 15 seconds, and wasn't putting anything in the stderr.txt about exiting (it already had threadsafe api at this time),
i remembered about the Collatz Cuda app's behavour and worked out that Boinc was killing the app at 10 seconds (or 15 seconds with Boinc 6.12.x and later) because it wasn't exiting when told to do so,
I managed to get Raistmer to do the exit check more often (I think he moved it from the outer loop to the middle loop), now app exits in well under a second, and doesn't cause a driver restart, the same could be done to the Collatz Cuda apps,

Claggy


Definition of insanity: Expecting "keep apps in memory" to terminate the application.

Would it really be that hard for BOINC not to lie to us and either not terminate apps when suspending or to change the text to say "keep CPU apps in memory" which is what it really does?

I guess I'm just in a lousy mood after spending hours and hours and hours trying to get the OpenCL app to perform as well as the CUDA or Brook+ apps. It looks like that just isn't going to happen.

Even worse, I installed CUDA 4.2 which required the 301.32 drivers. I made no changes to the source code but, after recompiling, it reports invalid results. Great. Nothing in the release notes hinted that any code I was using is no longer valid. Maybe I should just scrap the whole idea and deploy the 30% slower OpenCL app which does check much more often.

You know.... if everyone crunched Collatz and only Collatz all the time -- even when the computer is in use -- it would work fine the way it is. ;-)


Post to thread

Message boards : Windows : CUDA App - Not Threadsafe?


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.