No GPU WU?
log in

Advanced search

Message boards : Number crunching : No GPU WU?

Author Message
Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 201 - Posted: 28 Jul 2009, 8:10:52 UTC

Hello!

I can't get no GPU WU.
Although I have opt out the CPU version and only activated to get WUs for GPU, I keep getting CPU WUs no matter what I try.

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 213 - Posted: 28 Jul 2009, 12:04:53 UTC - in response to Message 201.

Hello!

I can't get no GPU WU.
Although I have opt out the CPU version and only activated to get WUs for GPU, I keep getting CPU WUs no matter what I try.


I assume you haven't unchecked the run GPU apps in your preferences since you were getting GPU WUs at one point.

You can try suspending all other CUDA app work on other projects and then update and see if that kick starts it. I've noticed some of the boinc versions have more trouble with cuda scheduling than others and I've got one box that refuses to download any CUDA work if there is work for any other CUDA app unless that other app is suspended.

If that doesn't work, my suggestion would be to try:
1. Set the project to no new work.
2. Abort the CPU tasks.
3. Reset the project. (If reset without aborting existing tasks, the WUs just go to giant bit-bucket in the sky and it takes 30 days for them to get re-issued.)
4. Set the project to allow new tasks.

If that doesn't work, turn on the sched_op_debug in the cc_config, and post what the log shows.

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 216 - Posted: 28 Jul 2009, 13:10:51 UTC - in response to Message 213.
Last modified: 28 Jul 2009, 13:37:12 UTC

I assume you haven't unchecked the run GPU apps in your preferences since you were getting GPU WUs at one point.

Always checked. I run GPUGrid normally.

You can try suspending all other CUDA app work on other projects and then update and see if that kick starts it.

Already tried, no help.

If that doesn't work, my suggestion would be to try:
1. Set the project to no new work.
2. Abort the CPU tasks.
3. Reset the project. (If reset without aborting existing tasks, the WUs just go to giant bit-bucket in the sky and it takes 30 days for them to get re-issued.)
4. Set the project to allow new tasks.

Tried that also some minutes ago. For a while I got no WUs at all, now I got CPU-WUs again, although I have them opt-out. :-\

I think the client version is the culprit, I use 6.6.20 for the moment and it already made some trouble on other projects as well. I will try another one now.


Edit:
Unbelievable, I now have the 6.6.38 and it is still the same!
Only GPU tasks allowed, but I still get none of them, but CPU tasks!

Btw:
The same problem is on AQUA with the server settings. But they have the option "if there's no work for the current app, give work from the other".
If the option is checked there, I get the right WUs, but not if this option is unchecked. Maybe it helps if you turn this feature on server-side.

frankhagen
Send message
Joined: 12 Jul 09
Posts: 188
Credit: 14,220,769
RAC: 1,424
Message 219 - Posted: 28 Jul 2009, 13:52:39 UTC - in response to Message 216.

Unbelievable, I now have the 6.6.38 and it is still the same!
Only GPU tasks allowed, but I still get none of them, but CPU tasks!


how many GPUgrid WU's do you have waiting in your queue?

____________

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 220 - Posted: 28 Jul 2009, 14:02:36 UTC - in response to Message 219.

how many GPUgrid WU's do you have waiting in your queue?

Only one I'm crunching on again now.
I don't think that is the reason that I can't get CUDA WUs here because on AQUA I can get some every time.

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 224 - Posted: 28 Jul 2009, 14:50:00 UTC
Last modified: 28 Jul 2009, 14:53:15 UTC

I've just turned on the scheduler_op_debug and this was what came out with GPUGrid on halt:
28.07.2009 16:42:50 Collatz Conjecture [sched_op_debug] CPU work request: 0.00 seconds; 0 idle CPUs
28.07.2009 16:42:50 Collatz Conjecture [sched_op_debug] CUDA work request: 518400.00 seconds; 1 idle GPUs
28.07.2009 16:42:55 Collatz Conjecture Scheduler request completed: got 0 new tasks
28.07.2009 16:42:55 Collatz Conjecture [sched_op_debug] Server version 607
28.07.2009 16:42:55 Collatz Conjecture Message from server: No work sent
28.07.2009 16:42:55 Collatz Conjecture Project requested delay of 121 seconds
28.07.2009 16:42:55 Collatz Conjecture [sched_op_debug] Deferring communication for 2 min 1 sec
28.07.2009 16:42:55 Collatz Conjecture [sched_op_debug] Reason: requested by project

I really suggest you're trying to turn on the "if there's no work for the selected app, send work from another" option for us.
At AQUA it only works with that option turned on, even if it sounds odd.

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 233 - Posted: 28 Jul 2009, 16:27:54 UTC - in response to Message 224.

I've just turned on the scheduler_op_debug and this was what came out with GPUGrid on halt:
28.07.2009 16:42:50 Collatz Conjecture [sched_op_debug] CPU work request: 0.00 seconds; 0 idle CPUs
28.07.2009 16:42:50 Collatz Conjecture [sched_op_debug] CUDA work request: 518400.00 seconds; 1 idle GPUs
28.07.2009 16:42:55 Collatz Conjecture Scheduler request completed: got 0 new tasks
28.07.2009 16:42:55 Collatz Conjecture [sched_op_debug] Server version 607
28.07.2009 16:42:55 Collatz Conjecture Message from server: No work sent
28.07.2009 16:42:55 Collatz Conjecture Project requested delay of 121 seconds
28.07.2009 16:42:55 Collatz Conjecture [sched_op_debug] Deferring communication for 2 min 1 sec
28.07.2009 16:42:55 Collatz Conjecture [sched_op_debug] Reason: requested by project

I really suggest you're trying to turn on the "if there's no work for the selected app, send work from another" option for us.
At AQUA it only works with that option turned on, even if it sounds odd.


There's only 1 application so there is no way it is giving work from another app. Aqua has two applications (AQUA and AQUA_CUDA) so it makes a difference there. Have you tried detaching and reattaching to the project? Also, what version of nVidia drivers are you using?

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 234 - Posted: 28 Jul 2009, 16:35:05 UTC - in response to Message 233.
Last modified: 28 Jul 2009, 16:38:27 UTC

Have you tried detaching and reattaching to the project? Also, what version of nVidia drivers are you using?

Yes of course, was one of the things I tried out as well. I've considered all possible problems which came to my mind, only thing I couldn't try was asking for work without any other GPU work active (or at stop), but I don't want to cancel my GPUGrid WU now, I wait until it is finished. If there's still work then here, I hope. ;-)
And my driver version is 185.38.

Btw:
I wonder...
If this is the continuation of 3x+1, why didn't you used the CPU apps from there?
Even the optimized version here takes too long, but the old apps on 3x+1 were mostly finished within 5 hours.

frankhagen
Send message
Joined: 12 Jul 09
Posts: 188
Credit: 14,220,769
RAC: 1,424
Message 235 - Posted: 28 Jul 2009, 16:41:23 UTC - in response to Message 234.

Even the optimized version here takes too long, but the old apps on 3x+1 were mostly finished within 5 hours.


even my slow 1.86 GHZ L-XEON finishes Collatz 1.1 in 7Ksecs - ???
____________

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 236 - Posted: 28 Jul 2009, 16:45:36 UTC - in response to Message 235.
Last modified: 28 Jul 2009, 16:46:01 UTC

even my slow 1.86 GHZ L-XEON finishes Collatz 1.1 in 7Ksecs - ???

That's strange.
I tried a WU this morning with the opt app and it only had about 3% after an hour on my AMD X2 5200.
Unfortunately it started over and over again, when it changed the project (forgot that I had the "leave apps in memory" deactivated), so I canceled it a while ago.
Maybe it does a huge jump later on?

Profile Gipsel
Volunteer moderator
Project developer
Project tester
Send message
Joined: 2 Jul 09
Posts: 279
Credit: 77,476,758
RAC: 76,461
Message 237 - Posted: 28 Jul 2009, 16:46:07 UTC - in response to Message 234.

I wonder...
If this is the continuation of 3x+1, why didn't you used the CPU apps from there?
Even the optimized version here takes too long, but the old apps on 3x+1 were mostly finished within 5 hours.

Have you looked at the computation times of the new CPU version? It is already significantly faster than the old 3x+1 application. And maybe I shouldn't say this at this stage, but most probably there is another round of major improvements for all versions (CPU and GPU) coming.

frankhagen
Send message
Joined: 12 Jul 09
Posts: 188
Credit: 14,220,769
RAC: 1,424
Message 238 - Posted: 28 Jul 2009, 16:50:31 UTC - in response to Message 236.

even my slow 1.86 GHZ L-XEON finishes Collatz 1.1 in 7Ksecs - ???

That's strange.
I tried a WU this morning with the opt app and it only had about 3% after an hour on my AMD X2 5200.
Unfortunately it started over and over again, when it changed the project (forgot that I had the "leave apps in memory" deactivated), so I canceled it a while ago.


CONGRATS to you and big thanks to DA again for his oh so fine working scheduler..

Maybe it does a huge jump later on?


it probably would have done..

____________

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 239 - Posted: 28 Jul 2009, 16:50:49 UTC - in response to Message 237.

Have you looked at the computation times of the new CPU version? It is already significantly faster than the old 3x+1 application. And maybe I shouldn't say this at this stage, but most probably there is another round of major improvements for all versions (CPU and GPU) coming.

See also my previous post.
Hm, okay, can't say much about it. I will try a CPU WU later on again with "leave apps" active. Maybe it really finishes faster than predicted.

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 254 - Posted: 28 Jul 2009, 19:22:13 UTC
Last modified: 28 Jul 2009, 19:32:52 UTC

Funny...
In order of trying something else I deleted the app_info.xml and the optimized app from the directory and let the client ask for work.
Guess what? I have now CUDA WUs! Can't believe it.
So the app_info.xml seems to has disturbed the whole progress, strange. Never thought that this could be the problem...

Edit:
Maybe it was the version discrepancy. Since the opt app is still 1.07 and the new apps 1.10 you may have to change something to use the opt app with CUDA as well.

Profile Cappy
Avatar
Send message
Joined: 12 Jul 09
Posts: 24
Credit: 18,485,830
RAC: 0
Message 255 - Posted: 28 Jul 2009, 19:37:37 UTC

ya i been smashing the cuda wu's since the ole slickman gave us work again


Post to thread

Message boards : Number crunching : No GPU WU?


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.