>BOINC V6.6.11&1/2 CUDA performance-prob found!

Questions and Answers : GPU applications : >BOINC V6.6.11&1/2 CUDA performance-prob found!
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 882607 - Posted: 5 Apr 2009, 21:08:06 UTC
Last modified: 5 Apr 2009, 21:27:08 UTC


I used BOINC V6.4.5 and .7 also V6.6.5 and .11 .
Everything fine.

I updated to V6.6.14 and .20 and saw at both 1/2 CUDA performance than before.

Well to see in TaskManager that the CUDA app get 1/2 CPU support.


I use Raistmer's V7 .dll's with CUDA V10 mod app.
And selfmade app_info.xml for only CUDA.


The prob couldn't be a mistake in my app_info.xml, or?

The 'old' CUDA V1.3 dll's of V7 mod?


What could be the problem?


Thanks a lot!



BTW.
I unchecked CPU usage for SETI@home. Only GPU usage.
But BOINC V6.6.5 to .20 ask for new work for the CPU.
I insert <avg_ncpus>2.00xx for CUDA [Quad-CPU + 2 GPUs rig] in the app_info.xml, that BOINC don't ask for 2 idle CPU-Core also.
But, with 'no idle CPUs', BOINC ask for new work for CPU and CUDA. [For CPU the double as for GPU]

ID: 882607 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 882623 - Posted: 5 Apr 2009, 22:01:55 UTC - in response to Message 882607.  

As soon as you add an app_info.xml file you are telling BOINC to forget about the preferences you set on the website and only use the settings in the anonymous platform file. So that is why you are trying to download for both CPU and GPU even though your preferences tell to do it for the GPU only. You will have to reflect that choice in your app_info.xml file, or delete the file and work with the preferences only.
ID: 882623 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 882629 - Posted: 5 Apr 2009, 22:26:45 UTC


What I could insert in the app_info.xml that BOINC don't ask for new CPU work?

ID: 882629 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 882632 - Posted: 5 Apr 2009, 22:45:39 UTC - in response to Message 882629.  

I don't know, you're not saying what you have in your app_info.xml at this moment. It's probably not something you have to add but more retract from the file. Like all hints towards CPU applications.
ID: 882632 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 882732 - Posted: 6 Apr 2009, 13:26:22 UTC
Last modified: 6 Apr 2009, 13:26:40 UTC


Some note to my 1/2 CUDA performance..

The 12/13.x credit WUs were running well with BOINC V6.6.14 and .20 .
But less performance I had with 38/42.x credit WUs.

What's different between this two AR-WUs, that I have 1/2 CUDA performance?


I will post later my app_info.xml..

I guess I have a mistake in my app_info.xml.
But this will make my less CUDA performance in V6.6.14 and .20?

ID: 882732 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 882987 - Posted: 7 Apr 2009, 3:38:50 UTC
Last modified: 7 Apr 2009, 3:52:57 UTC


O.K., here is my app_info.xml .

I would like to crunch only CUDA on my GPUs.
Nothing on the CPU.. only GPU support.. ;-)

How I could change my app_info.xml that BOINC don't ask for new work for the CPU?

And why I see 1/2 CUDA performance with BOINC V6.6.14 and .20 ?
Only with the 38/42.x credits WUs? [In the short test there were 12/13.x and 38/42.x credit WUs]

Thanks a lot!


<app_info>
<app>
<name>setiathome_enhanced</name>
</app>
<file_info>
<name>MB_6.08_mod_CUDA_V10.exe</name>
<executable/>
</file_info>
<file_info>
<name>cudart.dll</name>
<executable/>
</file_info>
<file_info>
<name>cufft.dll</name>
<executable/>
</file_info>
<file_info>
<name>libfftw3f-3-1-1a_upx.dll</name>
<executable/>
</file_info>
<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>608</version_num>
<flops>112000000000</flops>
<plan_class>cuda</plan_class>
<avg_ncpus>2.000000</avg_ncpus>
<coproc>
<type>CUDA</type>
<count>1</count>
</coproc>
<file_ref>
<file_name>MB_6.08_mod_CUDA_V10.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>cudart.dll</file_name>
</file_ref>
<file_ref>
<file_name>cufft.dll</file_name>
</file_ref>
<file_ref>
<file_name>libfftw3f-3-1-1a_upx.dll</file_name>
</file_ref>
</app_version>
</app_info>



The 112 GFLOPS are from the BOINC messages.
And very well if changing from V6.4.x to V6.6.x and back again.

ID: 882987 · Report as offensive
Claggy
Volunteer tester

Send message
Joined: 5 Jul 99
Posts: 4654
Credit: 47,537,079
RAC: 4
United Kingdom
Message 883694 - Posted: 9 Apr 2009, 14:22:49 UTC - in response to Message 882987.  
Last modified: 9 Apr 2009, 14:37:41 UTC

Here is an Cuda only app_info i posted about a week ago:

Re: V10 of modified SETI MB CUDA + opt AP package for full multi-GPU+CPU use

You shouldn't need to have avg_ncpus as 2.00, a figure of about 0.15 should be enough, I think you've seriously over estimating how much CPU is required to feed your GPU's,
Most of my RAC on my E8500 @ 4.14Ghz comes from doing Optimised Astropulse,
The flops count should be multipied by 0.2, the flops figure is used along side dcf to predict how much work is required and how long it'll take,
Too high a flops figure might mean work gets aborted before completion, because of maximum time exceeded.
Have you tried using the V10 .dll's i suggested you try?
Remember the later Boinc 6.6.x versions now show wall time instead of CPU time used.

From the Boinc Faq Service:

<ncpus>
Act as if there were N CPUs: run N tasks at once. This is for debugging, i.e. to simulate 2 CPUs on a machine that has only 1. Don't use it to limit the number of CPUs used by BOINC; use general preferences instead.

For use with CUDA: When you want to do CUDA only, not use the CPUs at all, set <ncpus>0</ncpus>
If you want to use all CPUs, either omit using this option, or set <ncpus>-1</ncpus>

Claggy
ID: 883694 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 884004 - Posted: 10 Apr 2009, 19:06:10 UTC


Thanks for reply!

So how I looked, your app_info.xml is similar to my app_info.xml .

So I'm confused, that I see as only one 1/2 CUDA performance > BOINC V6.6.11 .

My <avg_ncpus>2.000000 I use, because if only CUDA - BOINC would ask for work for 2 idle CPU-Core.
[Quad-CPU + 2 GPU]

ID: 884004 · Report as offensive
Claggy
Volunteer tester

Send message
Joined: 5 Jul 99
Posts: 4654
Credit: 47,537,079
RAC: 4
United Kingdom
Message 884018 - Posted: 10 Apr 2009, 19:42:04 UTC - in response to Message 884004.  


Thanks for reply!

So how I looked, your app_info.xml is similar to my app_info.xml .

So I'm confused, that I see as only one 1/2 CUDA performance > BOINC V6.6.11 .

My <avg_ncpus>2.000000 I use, because if only CUDA - BOINC would ask for work for 2 idle CPU-Core.
[Quad-CPU + 2 GPU]


Remember Boinc 6.6.11 is a development version, an Old one at that, Try Upgrading to Boinc 6.6.20, there's nine sets of bug fixes since then.

That's why you should use a cc_config with an <ncpus>0</ncpus> entry,
I'm using a <no_gpus>1</no_gpus> entry so my Laptop doesn't ask for GPU work since the GPU only has 128Mb, GPU now doesn't show up, and it only does CPU requests now.

Boinc Faq Service:

For use with CUDA: When you want to do CUDA only, not use the CPUs at all, set <ncpus>0</ncpus>
If you want to use all CPUs, either omit using this option, or set <ncpus>-1</ncpus>

Claggy
ID: 884018 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 885463 - Posted: 15 Apr 2009, 6:08:02 UTC
Last modified: 15 Apr 2009, 6:08:54 UTC


@ Claggy

Thanks a lot! :-)

Now with 'your' app_info.xml it's working fine with BOINC v6.6.20 .
I'm little confused why your .xml is 'better'.. ;-)
Because they are similar?!


@ all

O.K., but need to tell a BUG..

I reduced the CPUs in the computing preferences.
Simultaneously MB/CPU + CUDA/GPU.
I chose only use 2 CPUs - but 4 MBs/CPU were running.

If this because of the app_info.xml - I could limit the CPUs with <ncpus> in cc_config.xml ?


Only CUDA app_info.xml..

I unchecked CPU usage in the preferences. -> request for new CPU work, and 4 idle CPUs [every minute], [regardless also to GPU work request]....

Then additional..
I insert the <ncpus>0</ncpus> to cc_config.xml. -> request for new CPU work, and 4 idle CPUs [every minute], [regardless also to GPU work request] - and get answer "no work sent - check preferences" [this would mean high Berkeley server load!]

Then additional..
I made my <avg_ncpus>2.00x -trick in app_info.xml. -> request for new CPU work only simultaneously with GPU request [double CPU- as GPU- seconds], but get only CUDA work.. *happy*

[Quad-CPU + 2 GPUs - ONLY BOINC/SETI@home - no other project]

If I will finally insert the other 2 GPUs in the case, I don't need to mod the app_info.xml with <avg_ncpus> entry.. because without this entry I have '1.00 CPUs, 1 CUDA'.. so no idle CPU.. ;-)

ID: 885463 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 885621 - Posted: 15 Apr 2009, 18:59:45 UTC
Last modified: 15 Apr 2009, 19:07:33 UTC


Now I found the solution for 'my' problem! :-)

Woohoo.. :-D

I thought I'm stupid.. BUT - I'm not stupid.. I saw what I saw.. :-D


It wasn't my mistake with my app_info.xml..
See upper posts.. my was/is O.K. and similar to the which Claggy recommended..


BOINC was/is the problem! ;-D



Finally..
> BOINC V6.6.11 have problems to manage high cache with > ~ 2,500 WUs [~ 900 MB]!

AND because of this BOINC have 1/2 and more CUDA performance lost!



I made the update from BOINC V6.6.11 to V6.6.14 and V6.6.20 when I had a cache of ~ 3,700 WUs.
Woohoo..? No..! My rig make ~ 370 WUs/day.. a 10 day cache is finally ~ 3,700 WUs.


BOINC V6.4.5 and V6.4.7 and V6.6.5 and V6.6.11 have no problems to manage a cache of ~ 3,700 WUs.

WHY BOINC V6.6.14 and V6.6.20 which I tested??

What changed > BOINC V6.6.11 ?



You could say now.. why I need a soo big cache?

Soon I will insert two more GTX260 Core216 in the case..
Then the rig make ~ 740 WUs/day.. If I would stay with the current ~ 3,700 WUs/cache - then I would have 'only' 5 day/tasks..

I think very dangerous because of unplanned outages over loong weekends..


Now you could say.. take an other project as a 'backup/secure' project..
No thanks.. I don't 'waste' my electricity bill for an other project.. ONLY for my one and only loved project SETI@home! ;-D
Yes - I believe in SETI@home and want only support this!


It would be well to have a BOINC version which could manage ~ 7,400 or more tasks/cache..
I don't think only to my rig.. look to the top_host_list.. ;-)


It's possible to make real my dream?

ID: 885621 · Report as offensive
Profile zoom3+1=4
Volunteer tester
Avatar

Send message
Joined: 30 Nov 03
Posts: 66342
Credit: 55,293,173
RAC: 49
United States
Message 885639 - Posted: 15 Apr 2009, 19:49:08 UTC - in response to Message 885463.  


@ Claggy

Thanks a lot! :-)

Now with 'your' app_info.xml it's working fine with BOINC v6.6.20 .
I'm little confused why your .xml is 'better'.. ;-)
Because they are similar?!


@ all

O.K., but need to tell a BUG..

I reduced the CPUs in the computing preferences.
Simultaneously MB/CPU + CUDA/GPU.
I chose only use 2 CPUs - but 4 MBs/CPU were running.

If this because of the app_info.xml - I could limit the CPUs with <ncpus> in cc_config.xml ?


Only CUDA app_info.xml..

I unchecked CPU usage in the preferences. -> request for new CPU work, and 4 idle CPUs [every minute], [regardless also to GPU work request]....

Then additional..
I insert the <ncpus>0</ncpus> to cc_config.xml. -> request for new CPU work, and 4 idle CPUs [every minute], [regardless also to GPU work request] - and get answer "no work sent - check preferences" [this would mean high Berkeley server load!]

Then additional..
I made my <avg_ncpus>2.00x -trick in app_info.xml. -> request for new CPU work only simultaneously with GPU request [double CPU- as GPU- seconds], but get only CUDA work.. *happy*

[Quad-CPU + 2 GPUs - ONLY BOINC/SETI@home - no other project]

If I will finally insert the other 2 GPUs in the case, I don't need to mod the app_info.xml with <avg_ncpus> entry.. because without this entry I have '1.00 CPUs, 1 CUDA'.. so no idle CPU.. ;-)

No idle gpus or cpus here Sutaru. Of course this is for XP x64, But this could work in any Windows version that is supported by Seti and Boinc. And best of all no ncpus entries at all, But of course one needs Boinc 6.6.20 minimum.
Savoir-Faire is everywhere!
The T1 Trust, T1 Class 4-4-4-4 #5550, America's First HST

ID: 885639 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 885679 - Posted: 15 Apr 2009, 23:08:22 UTC - in response to Message 885639.  
Last modified: 15 Apr 2009, 23:25:53 UTC

No idle gpus or cpus here Sutaru. Of course this is for XP x64, But this could work in any Windows version that is supported by Seti and Boinc. And best of all no ncpus entries at all, But of course one needs Boinc 6.6.20 minimum.


I think you misread my posts.. ;-)


I want to have idle CPU-Core! :-)


AFAIK and from my experiences - only GPU calculation and the CPU only for support -> best GPU performance ! :-)


It depend what kind of rig.
When I'll find finally time to mod my case for 4 double GPUs.. then I'll have 1 GPU per 1 CPU-Core.
Nothing is better for pure GPU power! :-)
It would be better to have one CPU-Core more.. only for boinc.exe.. :-(


-----------------------------------------------------------------------
EDIT:
@ all
I edited/extended with my last upper post the title of this thread:
-prob found!

ID: 885679 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 886168 - Posted: 18 Apr 2009, 1:32:27 UTC - in response to Message 885621.  
Last modified: 18 Apr 2009, 1:33:27 UTC

...
> BOINC V6.6.11 have problems to manage high cache with > ~ 2,500 WUs [~ 900 MB]!

AND because of this BOINC have 1/2 and more CUDA performance lost!
...
...
BOINC V6.4.5 and V6.4.7 and V6.6.5 and V6.6.11 have no problems to manage a cache of ~ 3,700 WUs.

WHY BOINC V6.6.14 and V6.6.20 which I tested??

What changed > BOINC V6.6.11 ?
...


Someone sent my discovered BUG to the BOINC devs?


Or I need to post there:
http://boinc.berkeley.edu/dev


Or which way I need to go?


Now BOINC DEV-V6.6.23 is available. The prob is solved in this version?

ID: 886168 · Report as offensive
Profile Dirk Sadowski
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 886183 - Posted: 18 Apr 2009, 2:56:53 UTC
Last modified: 18 Apr 2009, 3:00:16 UTC


Maybe little note additional..

AFAIK - BOINC V6.4.x don't post the driver version to the Berkeley-overview-side.
BOINC V6.6.x do this.

But, after deinstallation of BOINC V6.6.x and latest .20 - BOINC V6.4.7 show now 'funny' driver versions.
For example:
driver: 925775673, 1667316016, -1, or others.. and sometimes nothing.. after every update a change..

What's broken in the deinsatallation prog?


EDIT:
If I remember correct - installed is nVIDIA driver V181.22.. :-)

ID: 886183 · Report as offensive

Questions and Answers : GPU applications : >BOINC V6.6.11&1/2 CUDA performance-prob found!


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.