Legacy Dual GPUs, BOINC assigned work to both but only one with load

\n studio-striking\n

Message boards : Number crunching : Legacy Dual GPUs, BOINC assigned work to both but only one with load
Message board moderation

To post messages, you must log in.

AuthorMessage
Bob

Send message
Joined: 4 Nov 17
Posts: 3
Credit: 8,800
RAC: 0
Message 7787 - Posted: 8 Nov 2017, 7:40:38 UTC

Hi I am running windows 7 64bit with BOINC 7.8.3 (x64) (was 7.8.2 at the time the images were taken)

I found the following problem 3 weeks ago when I swapped to Moo! wrapper but because I process as part of the GRCPool with gridcoin I could not post here as my user had no credit, last night I swapped one of the machines back to boincstat to get credit to post here. Mikey has tried to assist me via private messages so far. Thank you for that.

Problem described below.

I have 2 legacy ATI Radeon HD 3870 cards. When using Moo wrapper, BOINC assigns 2 tasks, 1 for each GPU. I used GPU-Z to investigate the GPUs when i noticed the one was running very hot compared to the other and that is when I discovered that only of the cards had any load. The one card runs 100% load while the other does nothing.

I have tried the following before posting.

1. Enable all GPUS in the BOINC config file.
2. Checked, mother board support PCIe x 16 2.0.to allow 2 cards to run at the same time - I confirmed mine does.
3. Plugged monitors into both cards - I did that.
4. Confirm that the 2nd card is not blown. I took them one by one and did a test - all good.
5. Make sure overdrive and crossfire is off - did that.

I logged this on the berkeley forum https://boinc.berkeley.edu/dev/forum_thread.php?id=11966&postid=82702#82702. The admins helped me to investigate and in the end their verdict was that this was a problem with the Moo wrapper and not with BOINC as BOINC was assigning the task to the correct GPU device and it appears as if the Moo wrapper was ignoring the BOINC instruction. From their posts I understand that there was an API change at that point that might have been missed.

Thanks for the assistance.

Wayne (Bob)

Here are some of the images.

BOINC shows tasks, one for each device 0 and device 1 - GPU-Z shows 0% Load on GPU device 0


GPU-Z shows 100% Load on GPU device 1


BOINC Eventlog


Machine summary from speccy


Speccy - GPU Device 0 details


Speccy - GPU Device 1 details


BOINC cc_config file.


BOINC App_config file (this makes no difference)


Taks manager - attempt to find dnetc wrapper commandline


init_data.xml file for device 0 Moo! tasks - shows BOINC has assigned the task to the correct GPU device.


init_data.xml file for device 1 Moo! task - shows BOINC has assigned the task to the correct GPU device.
ID: 7787 · Rating: 0 · rate: Rate + / Rate - Report as offensive
QuintLeo

Send message
Joined: 26 Jun 16
Posts: 53
Credit: 1,866,706,325
RAC: 0
Message 7790 - Posted: 9 Nov 2017, 2:28:45 UTC - in response to Message 7787.  

What is the "max concurrent" line for?
I've never had that in any of my multi-GPU systems.
ID: 7790 · Rating: 0 · rate: Rate + / Rate - Report as offensive
mmonnin

Send message
Joined: 29 Aug 16
Posts: 5
Credit: 682,585,026
RAC: 656,270
Message 7792 - Posted: 9 Nov 2017, 13:38:46 UTC
Last modified: 9 Nov 2017, 13:41:54 UTC

I had the same thing with NV GPUs like last week and just went to another project. Since They were not matched, a 1070 and 970, I didn't think they were being utilized the best. I tried the ngpus = 1 as well but the times did not line up. 22min for 1 task on both cards. 22 min for each task on d0 AND d1. One card should be faster than the other.

If I were to try again I was going to run another client and ignore the 1070 on one, and the 970 on the other to try and get separate tasks on each card.

max concurrent just limits the number of active tasks per client for that app. There is a project max concurrent as well. If you want to run 4 CPU threads of one project and 4 of another and force it to be that way, this line can be used instead of letting BM try to manage it. With 2 cards and it set to ngpu = 1 it's never enforced in this situation.
ID: 7792 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Bob

Send message
Joined: 4 Nov 17
Posts: 3
Credit: 8,800
RAC: 0
Message 7793 - Posted: 10 Nov 2017, 9:37:44 UTC - in response to Message 7792.  

I had the same thing with NV GPUs like last week and just went to another project. Since They were not matched, a 1070 and 970, I didn't think they were being utilized the best. I tried the ngpus = 1 as well but the times did not line up. 22min for 1 task on both cards. 22 min for each task on d0 AND d1. One card should be faster than the other.


So how does this work? Is there an admin here that feeds the bugs through to the developers?
ID: 7793 · Rating: 0 · rate: Rate + / Rate - Report as offensive
mikey
Avatar

Send message
Joined: 22 Jun 11
Posts: 2080
Credit: 1,842,737,672
RAC: 316,430
Message 7794 - Posted: 10 Nov 2017, 14:54:27 UTC - in response to Message 7792.  

I had the same thing with NV GPUs like last week and just went to another project. Since They were not matched, a 1070 and 970, I didn't think they were being utilized the best. I tried the ngpus = 1 as well but the times did not line up. 22min for 1 task on both cards. 22 min for each task on d0 AND d1. One card should be faster than the other.

If I were to try again I was going to run another client and ignore the 1070 on one, and the 970 on the other to try and get separate tasks on each card.


You can use a line in your cc_config.xml file to do that:

<exclude_gpu>
<url>http://boinc.fzk.de/poem/</url>
<device_num>1</device_num>
</exclude_gpu>

Those lines tell Boinc to exclude gpu 1 from the project poem, when Boinc first starts up in the event log tells you which gpu is 0 and which gpu is 1.
Put it in the <options> section like this:

<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
<exclude_gpu>
<url>http://boinc.fzk.de/poem/</url>
<device_num>1</device_num>
</exclude_gpu>
</options>
</cc_config>

To exclude your 2nd gpu from say Collatz you would add more lines:
<exclude_gpu>
<url>https://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
</exclude_gpu>

You put that just before or after the other gpu exclude lines.

Put together that would mean gpu 0 can run Poem but Poem but cannot run Collatz, and vice versa for gpu 1.

<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
</exclude_gpu>
<exclude_gpu>
<url>https://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
</exclude_gpu>
<exclude_gpu>
<url>http://boinc.fzk.de/poem/</url>
<device_num>1</device_num>
</options>
</cc_config>

This lets each gpu run it's own project with it's own priority settings.
ID: 7794 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Bob

Send message
Joined: 4 Nov 17
Posts: 3
Credit: 8,800
RAC: 0
Message 7798 - Posted: 11 Nov 2017, 9:50:47 UTC - in response to Message 7794.  

I had the same thing with NV GPUs like last week and just went to another project. Since They were not matched, a 1070 and 970, I didn't think they were being utilized the best. I tried the ngpus = 1 as well but the times did not line up. 22min for 1 task on both cards. 22 min for each task on d0 AND d1. One card should be faster than the other.

If I were to try again I was going to run another client and ignore the 1070 on one, and the 970 on the other to try and get separate tasks on each card.


You can use a line in your cc_config.xml file to do that:

<exclude_gpu>
<url>http://boinc.fzk.de/poem/</url>
<device_num>1</device_num>
</exclude_gpu>

Those lines tell Boinc to exclude gpu 1 from the project poem, when Boinc first starts up in the event log tells you which gpu is 0 and which gpu is 1.
Put it in the <options> section like this:

<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
<exclude_gpu>
<url>http://boinc.fzk.de/poem/</url>
<device_num>1</device_num>
</exclude_gpu>
</options>
</cc_config>

To exclude your 2nd gpu from say Collatz you would add more lines:
<exclude_gpu>
<url>https://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
</exclude_gpu>

You put that just before or after the other gpu exclude lines.

Put together that would mean gpu 0 can run Poem but Poem but cannot run Collatz, and vice versa for gpu 1.

<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
</exclude_gpu>
<exclude_gpu>
<url>https://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
</exclude_gpu>
<exclude_gpu>
<url>http://boinc.fzk.de/poem/</url>
<device_num>1</device_num>
</options>
</cc_config>

This lets each gpu run it's own project with it's own priority settings.


Thanks I have now done the same till this bug is fixed. Moo on one GPU and something else on the other.
ID: 7798 · Rating: 0 · rate: Rate + / Rate - Report as offensive
Profile [AF>EDLS]GuL

Send message
Joined: 11 May 17
Posts: 2
Credit: 45,017,134
RAC: 708
Message 7846 - Posted: 15 Dec 2017, 21:37:27 UTC - in response to Message 7787.  

I have 2 legacy ATI Radeon HD 3870 cards. When using Moo wrapper, BOINC assigns 2 tasks, 1 for each GPU. I used GPU-Z to investigate the GPUs when i noticed the one was running very hot compared to the other and that is when I discovered that only of the cards had any load. The one card runs 100% load while the other does nothing.


Hello,
I have a very similar problem : with Dnetc 1.04 and AMD application, it is impossible to compute on AMD GPU. The application is rather using the integrated GPU (Intel HD Graphics 4600 on a Core i7 4790K). I have tested separately with two GPU : a HD 7970 and a R9 Fury.

It is the same that was suggested there: https://moowrap.net/forum_thread.php?id=484&postid=7674#7674

I have tried to update the amd driver, and to blacklist the gpu, without success. Could it be a command line switch to add in an app_config.xml file, to select the correct gpu ?

Thanks
ID: 7846 · Rating: 0 · rate: Rate + / Rate - Report as offensive
QuintLeo

Send message
Joined: 26 Jun 16
Posts: 53
Credit: 1,866,706,325
RAC: 0
Message 7873 - Posted: 9 Jan 2018, 23:40:13 UTC - in response to Message 7846.  

I believe there is an "ignore intel" option among the GPU ignore commands?
That's an option in config.xml though not app_config.xml
It's wierd that the client would ONLY work on the Intel iGPU though - it should be using both the Intel AND the AMD GPUs in the system if you have allgpus set.
ID: 7873 · Rating: 0 · rate: Rate + / Rate - Report as offensive
mikey
Avatar

Send message
Joined: 22 Jun 11
Posts: 2080
Credit: 1,842,737,672
RAC: 316,430
Message 7876 - Posted: 10 Jan 2018, 12:40:39 UTC - in response to Message 7846.  

I have 2 legacy ATI Radeon HD 3870 cards. When using Moo wrapper, BOINC assigns 2 tasks, 1 for each GPU. I used GPU-Z to investigate the GPUs when i noticed the one was running very hot compared to the other and that is when I discovered that only of the cards had any load. The one card runs 100% load while the other does nothing.


Hello,
I have a very similar problem : with Dnetc 1.04 and AMD application, it is impossible to compute on AMD GPU. The application is rather using the integrated GPU (Intel HD Graphics 4600 on a Core i7 4790K). I have tested separately with two GPU : a HD 7970 and a R9 Fury.

It is the same that was suggested there: https://moowrap.net/forum_thread.php?id=484&postid=7674#7674

I have tried to update the amd driver, and to blacklist the gpu, without success. Could it be a command line switch to add in an app_config.xml file, to select the correct gpu ?

Thanks


Go to Preferences under Your Account here on the website and uncheck the 'use intel' box:
Resource share 100 --- --- ---
Use CPU --- --- ---
Use ATI GPU --- --- ---
Use NVIDIA GPU
ID: 7876 · Rating: 0 · rate: Rate + / Rate - Report as offensive

Message boards : Number crunching : Legacy Dual GPUs, BOINC assigned work to both but only one with load


 
Copyright © 2011-2024 Moo! Wrapper Project