V10/11 of modified SETI MB CUDA + opt AP package for full multi-GPU+CPU use

[Kashra]

Hey,

have you tried to run one project solo? You can turn off AP in your Boinc main account settings in setiathome.de
perhaps the AP code is messing around with the Seti apps especially since you have a i7 920 cpu with the "hyperthreading" feature. Could really be a bottleneck caused by I/O from AP or seti running together. Just try to run seti alone and recheck your times, if they are fine you got your solution - use the cc_config file to limit Seti+AP e.g. to 9 or 10 procs including the GPUs to have 1 or 2 CPU cores left for system and kernel stuff.

[codeman05]

I was thinking it could be a I/O issue as well.

How exactly do I limit the number of active CPU's in the cc_config.
Is it still the: <ncpus>6</ncpus> for example?

UPDATE:
I just dropped the Processor Utilization in the BOINC pref's to 50%, so now I have 4 AP's running and 3 CUDAs......CUDA WU time dropped from 17min to 10 mins...
I think you were right.

I guess the real question then is, is the "loss" of 4 AP units worth the 25-50% decrease in CUDA WU times.

[Raistmer]

I just dropped the Processor Utilization in the BOINC pref's to 50%, so now I have 4 AP's running and 3 CUDAs......CUDA WU time dropped from 17min to 10 mins...

Very interesting.
I did some standalone benchmarking of CUDA MB while CPU running 4 another BOINC tasks (not AP though, einstein@home ones) and didn't see any degradation in CUDA MB elapsed times and speed.
But it was quad w/o HT.
Do you use V10 or V11 CUDA MB build from this thread or stock one?

[Raistmer]

I guess the real question then is, is the "loss" of 4 AP units worth the 25-50% decrease in CUDA WU times.

Well, if forget about fact that 1 sec of AP work gives better credit than 1 sec of SETI MB work it seems you didn't change total host performance.
There were reports that with HT on 8 AP took twice the time 4 AP take with HT off on the same host. You leaved HT on but affectively disable some on CPU cores.
Try to compute overall host work done in both your configs (unfortunately, credits not very suitable for that cause AP and MB credit rates different).

[Raistmer]

New VLAR limit value used (0.12 instead of 0.14). For details see this thread:
http://lunatics.kwsn.net/12-gpu-crunching/gpu-cpu-pefrmance-dependence-from-ar-value.msg16800.html#msg16800

[attachment deleted by admin]

[codeman05]

I just dropped the Processor Utilization in the BOINC pref's to 50%, so now I have 4 AP's running and 3 CUDAs......CUDA WU time dropped from 17min to 10 mins...

Very interesting.
I did some standalone benchmarking of CUDA MB while CPU running 4 another BOINC tasks (not AP though, einstein@home ones) and didn't see any degradation in CUDA MB elapsed times and speed.
But it was quad w/o HT.
Do you use V10 or V11 CUDA MB build from this thread or stock one?

I'm using the V11 that was referenced a few pages back. I'm now had a full 24 hours of crunching with HT more or less disabled (thru BOINC anyway). AP processing time dropped from about 17-20 hours down to 10-11 hours consistant. CUDA WU's are still pumping out in 10 minutes solid compared with 15-20 mins prior.

This seems to confirm what you were saying about HT rigs crunching slower, which makes since since the four HT "cores" are not true physical CPUs.
So far it would seem that my overall production has increased. Due to the decrease in time to process AP units, I'm not that much slower on AP than before, and cuda is running close to twice as fast....not bad.

[rich1ell]

I wonder if someone could help.

I've got a quad 6600 with a 9600gt cuda gpu running Win XP pro with bionc 6.6.20.
I'm wanting to run AP wu's using an optimised app on all 4 cpu cores and stock (or opt i'm not bothered) app for MB608 wu's on my GPU.

I've tried "Raistmer's_opt_pack_V10a_x86_SSSE3_MB_CPU_GPU_team" pack but when I restart bionc it drops all my cuda wu's but processes my AP wu's great. This is what I've been doing:

• Stop network activity in bionc
• Close bionc and check processes have stopped
• Back up my bionc dir
• Drop opt V10 apps in to seti data directory - incuding packaged app_info file
• "number_of_GPUs" file is '1'
• cc_config = <ncpus>5</ncpus> (4 cores and 1 gpu)
• Restart bionc

But its annoying how I cant get it to run cuda especially as everything works great, all be it slow, with just the stock apps from seti.
cheers Rich

[rich1ell]

I wonder if someone could help.

I've got a quad 6600 with a 9600gt cuda gpu running Win XP pro with bionc 6.6.20.
I'm wanting to run AP wu's using an optimised app on all 4 cpu cores and stock (or opt i'm not bothered) app for MB608 wu's on my GPU.

I've tried "Raistmer's_opt_pack_V10a_x86_SSSE3_MB_CPU_GPU_team" pack but when I restart bionc it drops all my cuda wu's but processes my AP wu's great. This is what I've been doing:

• Stop network activity in bionc
• Close bionc and check processes have stopped
• Back up my bionc dir
• Drop opt V10 apps in to seti data directory - incuding packaged app_info file
• "number_of_GPUs" file is '1'
• cc_config = <ncpus>5</ncpus> (4 cores and 1 gpu)
• Restart bionc

But its annoying how I cant get it to run cuda especially as everything works great, all be it slow, with just the stock apps from seti.
cheers Rich

In addition this is my host: http://setiathome.berkeley.edu/show_host_detail.php?hostid=4819034

I have just tried again with the V10 and updated my nvidia drivers to 182.50 and haverecorded the error message in bionc  as:

19/04/2009 17:27:29   SETI@home   [error] No app version for result: windows_intelx86 608 cuda

This message occurs for each cuda I have in the cache. all AP restarted fine
Rich

[Raistmer]

check your app_info file. It should contain corret app name (the same you have in project folder)

[rich1ell]

check your app_info file. It should contain corret app name (the same you have in project folder)

Everything seems to be in order with the app_info as far as I can tell, here is the text from the file:

<app_info>
    <app>
        <name>astropulse</name>
    </app>
    <file_info>
        <name>ap_5.00r103_SSE3.exe</name>
        <executable/>
    </file_info>
    <app_version>
        <app_name>astropulse</app_name>
        <version_num>500</version_num>
        <file_ref>
            <file_name>ap_5.00r103_SSE3.exe</file_name>
            <main_program/>
        </file_ref>
    </app_version>
   <app>
      <name>astropulse_v5</name>
   </app>
   <file_info>
           <name>ap_5.03r112_SSE3.exe</name>
           <executable/>
       </file_info>
       <app_version>
           <app_name>astropulse_v5</app_name>
           <version_num>503</version_num>
           <file_ref>
                  <file_name>ap_5.03r112_SSE3.exe</file_name>
                  <main_program/>
           </file_ref>
       </app_version>
<app>
<name>setiathome_enhanced</name>
</app>
<file_info>
<name>AK_v8b_win_SSSE3x_GPU_CPU_team_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>
<file_info>
<name>MB_6.08_mod_CUDA_V10.exe</name>
<executable/>
</file_info>

<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>528</version_num>
<file_ref>
<file_name>AK_v8b_win_SSSE3x_GPU_CPU_team_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>
<file_ref>
<file_name>MB_6.08_mod_CUDA_V10.exe</file_name>
</file_ref>
</app_version>

<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>603</version_num>
<file_ref>
<file_name>AK_v8b_win_SSSE3x_GPU_CPU_team_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>
<file_ref>
<file_name>MB_6.08_mod_CUDA_V10.exe</file_name>
</file_ref>
</app_version>


<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>607</version_num>
<file_ref>
<file_name>AK_v8b_win_SSSE3x_GPU_CPU_team_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>
<file_ref>
<file_name>MB_6.08_mod_CUDA_V10.exe</file_name>
</file_ref>
</app_version>

<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>608</version_num>
<file_ref>
<file_name>AK_v8b_win_SSSE3x_GPU_CPU_team_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>
<file_ref>
<file_name>MB_6.08_mod_CUDA_V10.exe</file_name>
</file_ref>
</app_version>

</app_info>

[Raistmer]

<core_client_version>6.6.20</core_client_version>
<![CDATA[
<message>
CreateProcess() failed - </message>
]]>
"teamed" AK_v8 app can't spawn CUDA MB app possibly. Check value in Number_of_GPUs file.
Also, check if you use BOINC service install. Maybe BOINC "protected" install short of spawn child process access right? ...

[Slagathor]

Got it working finally thanks to Claggy, but I do have a question.

It isnt a really big deal, but just curious why the wu times would increase?? They went up about 5 minutes from the stock one.

From 17-18 minutes to 22-23 minutes per.

[Claggy]

Got it working finally thanks to Claggy, but I do have a question.

It isnt a really big deal, but just curious why the wu times would increase?? They went up about 5 minutes from the stock one.

From 17-18 minutes to 22-23 minutes per.

Is that while doing Astropulse at the same time?, I Know the version of Boinc (6.6.23) i'm running shows Elapsed time as apposed to CPU time now,
I can't say I've seen Cuda WU's generally taking longer, unless you've now got WU's with a High AR instead of a Mid Range AR,
Astropulse WU's are a different matter, Boinc shows elapsed time, but reports CPU time,
and while I'm doing Cuda WU's there can be an hour's differerence between the two.

Claggy

[Slagathor]

Got it working finally thanks to Claggy, but I do have a question.

It isnt a really big deal, but just curious why the wu times would increase?? They went up about 5 minutes from the stock one.

From 17-18 minutes to 22-23 minutes per.

Is that while doing Astropulse at the same time?, I Know the version of Boinc (6.6.23) i'm running shows Elapsed time as apposed to CPU time now,
I can't say I've seen Cuda WU's generally taking longer, unless you've now got WU's with a High AR instead of a Mid Range AR,
Astropulse WU's are a different matter, Boinc shows elapsed time, but reports CPU time,
and while I'm doing Cuda WU's there can be an hour's differerence between the two.

Claggy

All that rig does it crunch CUDA's, no AP's on that one, it is using 6.6.20.

[Raistmer]

Please, before asking "why my crunch time increased/decreased" check if you compare times for results with THE SAME Angle Range (AR) value.
Time for crunching tasks with different ARs differ strongly both for CPU and GPU apps. It's meaningless to compare results with different ARs to conclude that your host become suddenly faster or slower.