Confirmation errors
log in

Advanced search

Message boards : Windows : Confirmation errors

Author Message
Andi-Hagen
Send message
Joined: 6 Jan 12
Posts: 10
Credit: 158,051,619
RAC: 121,835
Message 21306 - Posted: 19 Sep 2015, 9:56:05 UTC

Since I have changed from 6.04 to 6.06 (CUDA55) I get errors for my computed tasks. The status in my task list is "Bestätigungsfehler" (confirmation error).
Task ids for example: 23016670, 23003597, 22975627.
I am using a NVIDIA GeForce GTX 660 (2048 MB) with Windows 7 SP1.

My cuda55.config is:
verbose=1
threads=8
items_per_kernel=18
kernels_per_reduction=8
sleep=1
lut_size=19

What is going wrong?

And it has become slower. Solo Collatz with 6.04 took about 60 minutes, now it takes 80 minutes.

Profile Thomas Krajewsky
Avatar
Send message
Joined: 2 Feb 10
Posts: 13
Credit: 379,296,674
RAC: 0
Message 21307 - Posted: 19 Sep 2015, 15:42:56 UTC - in response to Message 21306.

Since I have changed from 6.04 to 6.06 (CUDA55) I get errors for my computed tasks. The status in my task list is "Bestätigungsfehler" (confirmation error).
Task ids for example: 23016670, 23003597, 22975627.
I am using a NVIDIA GeForce GTX 660 (2048 MB) with Windows 7 SP1.

My cuda55.config is:
verbose=1
threads=8
items_per_kernel=18
kernels_per_reduction=8
sleep=1
lut_size=19

What is going wrong?

And it has become slower. Solo Collatz with 6.04 took about 60 minutes, now it takes 80 minutes.


lut_size = 19 gives 8*2^19=4194304, that's 4MB, the card has only a cache of 384KB

For better values in the .config:
http://boinc.thesonntags.com/collatz/forum_thread.php?id=1009&postid=16503#16503

GeForce GTX 660, Keppler:
960 CUDA cores
384 KB Cache
2048 MB RAM

Try this
verbose=1
items_per_kernel=20
kernels_per_reduction=9
threads=9
sleep=1
lut_size=15


Or try the sieve (beta) apps, they run now very well without validation errors.

Andi-Hagen
Send message
Joined: 6 Jan 12
Posts: 10
Credit: 158,051,619
RAC: 121,835
Message 21313 - Posted: 20 Sep 2015, 15:48:15 UTC - in response to Message 21307.

Since I have changed from 6.04 to 6.06 (CUDA55) I get errors for my computed tasks. The status in my task list is "Bestätigungsfehler" (confirmation error).
Task ids for example: 23016670, 23003597, 22975627.
I am using a NVIDIA GeForce GTX 660 (2048 MB) with Windows 7 SP1.

My cuda55.config is:
verbose=1
threads=8
items_per_kernel=18
kernels_per_reduction=8
sleep=1
lut_size=19

What is going wrong?

And it has become slower. Solo Collatz with 6.04 took about 60 minutes, now it takes 80 minutes.


lut_size = 19 gives 8*2^19=4194304, that's 4MB, the card has only a cache of 384KB

For better values in the .config:
http://boinc.thesonntags.com/collatz/forum_thread.php?id=1009&postid=16503#16503

GeForce GTX 660, Keppler:
960 CUDA cores
384 KB Cache
2048 MB RAM

Try this
verbose=1
items_per_kernel=20
kernels_per_reduction=9
threads=9
sleep=1
lut_size=15


Or try the sieve (beta) apps, they run now very well without validation errors.


I tried with your config parameters. The GPU-Load raises to 99%, it is impossible to work with the computer in parallel and the runtime raises up to 100 minutes for a Solo Collatz.
I tried the sieve apps. All jobs fail with computation error (task-ids: 23434614, 23434613, 23434610)
Can I go back to 6.04? I was quite happy with that version.

Andi-Hagen
Send message
Joined: 6 Jan 12
Posts: 10
Credit: 158,051,619
RAC: 121,835
Message 21314 - Posted: 20 Sep 2015, 17:40:11 UTC - in response to Message 21313.

I forgot to mention that I still get confirmation errors.
Task-IDs: 23434598, 23434596.

Andi-Hagen
Send message
Joined: 6 Jan 12
Posts: 10
Credit: 158,051,619
RAC: 121,835
Message 21315 - Posted: 20 Sep 2015, 17:42:28 UTC - in response to Message 21314.

Sorry, wrong task IDs. I got cofirmations errors for 23155257 and 23070558.

Andi-Hagen
Send message
Joined: 6 Jan 12
Posts: 10
Credit: 158,051,619
RAC: 121,835
Message 21316 - Posted: 20 Sep 2015, 17:54:13 UTC - in response to Message 21315.

I found the hint about VS2012 C++ runtime update. I have installed SP4 and now the Sieve apps are running.

Profile Thomas Krajewsky
Avatar
Send message
Joined: 2 Feb 10
Posts: 13
Credit: 379,296,674
RAC: 0
Message 21317 - Posted: 20 Sep 2015, 18:56:50 UTC - in response to Message 21316.

I found the hint about VS2012 C++ runtime update. I have installed SP4 and now the Sieve apps are running.


That's fine, here my .config for sieve wus with GTX 660:
verbose=1
items_per_kernel=20
kernels_per_reduction=48
threads=10
sleep=1
lut_size=15
reduceCPU=0
sieve_size=30

(The values items_per_kernel and reducecpu are currently with V1.21 not supported.)

If your computer gets unusuable, try increasing the sleep value and/or decreasing the other values to your old working ones.

Another solution maybe an option in boinc manager, you can tell it to stop GPU processing for x minutes if you use keyboard or mouse.

Third possible solution: You may use two sets of .config files, one for working, the other for crunching and two batch files for copying the appropiate files into boincs data directory.

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 21319 - Posted: 21 Sep 2015, 4:39:06 UTC - in response to Message 21317.

I found the hint about VS2012 C++ runtime update. I have installed SP4 and now the Sieve apps are running.


That's fine, here my .config for sieve wus with GTX 660:
verbose=1
items_per_kernel=20
kernels_per_reduction=48
threads=10
sleep=1
lut_size=15
reduceCPU=0
sieve_size=30

(The values items_per_kernel and reducecpu are currently with V1.21 not supported.)

If your computer gets unusuable, try increasing the sleep value and/or decreasing the other values to your old working ones.

Another solution maybe an option in boinc manager, you can tell it to stop GPU processing for x minutes if you use keyboard or mouse.

Third possible solution: You may use two sets of .config files, one for working, the other for crunching and two batch files for copying the appropiate files into boincs data directory.


sieve_size=26 means 2^26 numbers per kernel. using 30 means 2^30 numbers per kernel of 16 times more than the default. Some GPUs may be able to use 30 but it will likely cause the driver to crash on most GPUs. That's why v1.04 is not the current sieve app. It used 2^30 items per kernel which was too many resources for most GPUs, especially AMD GPUs. That's because it will take more than 33 milliseconds to complete the kernel and Windows assumes that anything over 33 milliseconds means that the driver has crashed and will reset it.

FYI, reduce_cpu=0 is the default in the sieve v1.21 app although you can still set reduce_cpu=1 which will use the CPU to to the summarization but the larger the sieve_size setting, the more work the CPU will then have to do. The items_per_kernel parameter is not supported because sieve_size now controls the items per kernel.

Profile Thomas Krajewsky
Avatar
Send message
Joined: 2 Feb 10
Posts: 13
Credit: 379,296,674
RAC: 0
Message 21321 - Posted: 21 Sep 2015, 7:38:16 UTC - in response to Message 21319.

FYI, reduce_cpu=0 is the default in the sieve v1.21 app although you can still set reduce_cpu=1 which will use the CPU to to the summarization but the larger the sieve_size setting, the more work the CPU will then have to do. The items_per_kernel parameter is not supported because sieve_size now controls the items per kernel.

reduce_cpu with '_' is the correct version?
I started without '_', i copied that from your post:
http://boinc.thesonntags.com/collatz/forum_thread.php?id=1273&postid=20542#20542

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 21361 - Posted: 30 Sep 2015, 4:03:17 UTC - in response to Message 21321.

FYI, reduce_cpu=0 is the default in the sieve v1.21 app although you can still set reduce_cpu=1 which will use the CPU to to the summarization but the larger the sieve_size setting, the more work the CPU will then have to do. The items_per_kernel parameter is not supported because sieve_size now controls the items per kernel.

reduce_cpu with '_' is the correct version?
I started without '_', i copied that from your post:
http://boinc.thesonntags.com/collatz/forum_thread.php?id=1273&postid=20542#20542


Somewhere along the line, I changed it to be consistent with the underscores in sieve_size, lut_size, and kernels_per_iteration. I've updated the other thread.


Post to thread

Message boards : Windows : Confirmation errors


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.