Seti@Home optimized science apps and information
Optimized Seti@Home apps => Windows => GPU crunching => Topic started by: FalconFly on 10 May 2009, 09:35:01 am
-
Following the Team Mod discussion and updates, I went with the new recommendations.
However, the old result remained :
BOINC 6.6.20 x64 does not recognize the 2nd GPU of my GTX295 unter Win7 64bit.
Consequently the V11 VLARkill_refined is launched only once instead of twice.
Other than that, everthing is apparently working fine.
I'll have to revert to my old Installation with V10a Team Mod in order to get the 2nd GPU back into business unless someone can find an error I made.
- Point blank installation of BOINC x64 6.6.20
- attached to Project (no workunits downloaded yet)
- quite BOINC
- inserted cudart.dll into program files Directory (overwriting the default one, assuming the new one is required here also)
- inserted V11 + SSE3 Files, cudart.dll, cufft.dll, libfftw3f-3-1-1a_upx.dll and app_info.xml files into program data/projects/setiathome Sub-Directory
- launched BOINC
-- edit --
This app_info.xml works like a charm for my 2 GTX295, now running under WinXP SP3 32bit and BOINC 6.6.20 Win32
(Win32, CPU SSE3 for AP und MB, for GPU V11 VLARkill)
<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_v8_win_SSE3.exe</name>
<executable/>
</file_info>
<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>528</version_num>
<file_ref>
<file_name>AK_v8_win_SSE3.exe</file_name>
<main_program/>
</file_ref>
</app_version>
<app_version>
<app_name>setiathome_enhanced</app_name>
<version_num>603</version_num>
<file_ref>
<file_name>AK_v8_win_SSE3.exe</file_name>
<main_program/>
</file_ref>
</app_version>
<app>
<name>setiathome_enhanced</name>
</app>
<file_info>
<name>MB_6.08_mod_CUDA_V11_VLARKill_refined.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>21400000000</flops>
<plan_class>cuda</plan_class>
<avg_ncpus>0.150000</avg_ncpus>
<max_ncpus>0.150000</max_ncpus>
<coproc>
<type>CUDA</type>
<count>1</count>
</coproc>
<file_ref>
<file_name>MB_6.08_mod_CUDA_V11_VLARKill_refined.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>
-
Hi FalconFly,
Was curious about this issue you're having, in case it has ramifications for some of the development going on in the background, regarding installer prototype:
Number of GPUs is managed by Boinc & Drivers:
i notice that your Windows 7 machine, comes up with coprocessor as:
NVIDIA GeForce GTX 295 (896MB) driver: 18585
Whereas Bob Mahoney's 3 card (6 GPU) setup comes up as:
[6] NVIDIA GeForce GTX 295 (895MB) driver: 18585
Just to eliminate the obvious things that can cause this: Do you have SLI disabled properly (so it'll be seen by the system as 2 GPUs instead of one big one), and does the seond GPU possibly need something plugged into it to wake it up. The v10 team build may be 'Fudging' the setup to make it look like it's working, when it might not be.
Jason
-
Yes, that's correct.
I've never managed to get 6.6.20 to show [2] GPUs like I saw on other Systems I looked at for comparison.
I have SLI disabled in the Driver for the GTX295 and even Enabled it for a quick test. No change.
Unless I made another mistake this as good as I could get, following the latest recommendations.
Lacking detailled documentation (I hate that), there still are possible mistakes I did :
Is the cudart.dll indeed supposed not only to go into the Projects directory, but also in the program files?
The one I found there by default was a ~447k File and came with the original BOINC 6.6.20 installation.
I noticed the one found in the V10a was only 228k and thus obviously another version.
I haven't checked so far, but maybe the cudart.dll coming with BOINC 6.6.20 Installation by default might/should be used instead of the one I took from the V10a Team Mod package.
Might also be an Idea to let BOINC download its .dlls from the stock CUDA application download alltogether and try to use these instead with the V11 VLARkill_refined (and leave all .dlls from the old V10a Team Mod package alone ?
With a bit of luck, one of these steps I took turns out to be a mistake.
My other Single GPU GTX285 installation with BOINC 6.6.20 does run fine this way, but I also noted there ist no [1] in front of the GPU indication on the Computer Overview page. With these details, we may be onto something.
Additionally, it runs the Default intel_x86 CUDA client which still leaves me with the VLAR problems.
-
it seems that is a driver bug - i have three cards and see only onr too under win7...
i am trying another driver - if functional then i refer and send a link ...
cuda-z is showing only one cuda device too ....
-
-- edit --
http://setiathome.berkeley.edu/show_host_detail.php?hostid=4900487
Same Driver, so it does seem to work under Windows 7 x64 at all, but these are 2 individual GTX285...
Might be a problem with the dual GPU cards then.
-
-- edit --
http://setiathome.berkeley.edu/show_host_detail.php?hostid=4900487
Same Driver, so it does seem to work under Windows 7 x64 at all, but these are 2 individual GTX285...
Might be a problem with the dual GPU cards then.
Win2003 Server x64 works OK with 182.50 and 2 GPUs.
You could try this OS or WinXP 64.
It seems new M$ OSes should be avoided as long as possible if you value your time and nerves ;)
-
Hehe, I hear you ;)
The horrors I went through just navigating through Windows 7 RC were insane, even with everthing possible switched back to "oldschool" I can already say that I will avoid that for as long as possible.
(since it's supposed to run just less than a month for me and it was free, I just took a shot at it)
I sure wish there would be a Linux CUDA SETI available already, then I'd happily take that.
-
now i am using two gpus - one as primary display and second as psychx card ...
-
That is a problem identified by Paul Buck BOINC and one of the reasons that even into 6.6.23. The logic for detecting of GPU's had a problem, the fix disabled Some GPU's because they were not similar. Paul's 295 (times 2 for 4 GPU's) reported less memory for the second GPU, thus were negated. If I recall correctly, it was corrected in 6.6.26 but then other issues were reported. 6.6.28 is the current testing version which has proper detection working.
-
Ha !
IT WORKS ! :) :)
Definitely the combination of Windows 7 x64, Dual GPU card (GTX295) and the x64 185.85WHQL Driver that caused my Problems.
Now BOINC 6.6.20 Win32 runs under XP SP3 32bit, and immediately recognized all 4 GPUs of my two GTX295, my app_info.xml works like a charm there (updated my previous post).
So it looks like Single GPU cards will likely run normal in pairs/triple even under Windows 7 (such Host exists in the SETI Top Computer ranking), but it fails to do so when DualGPU Cards (like the GTX295) are used.
I assume another Driver (other than 185.85WHQL) might fix the problem, otherwise the V10a Team Mod is the only way I know to fill DualGPU cards under Windows 7 at this time...
-
lol
Not realizing two GTX295 would already steal some 3.5GB of upper memory, the System crashed hard and I had to remove one of the GTX295 again. With only some ~400MB usable RAM remaining, operation of the GTX295 pair is basically useless under a 32bit OS. I simply forgot that.
So right now I'm down to one GTX295 per System as it persistently failed unter WinXP 32 (insufficient address space), Vista 64 (Driver bug) and Win7 64 (Driver bug).
I've tried plugging in multiple Monitors as some suggested in the Forums, but that didn't change anything either.
It is beyond me how people manage to operate more than one GTX295 with the 185.85 Driver and due to time I've given up on that project. I've tried hard for a whole week but all to no avail.
-
You could try to use other GPU for PhysX instead of plugging monitor. This officially recommended from nVidia (worth to look their forums btw for GPU issues).
-
I tried enabling Physx but that did not change anything unfortunately.
The only working Config I found for 2xGTX295 was WinXP32 (couldn't work due to RAM limitations, but otherwise would have).
I'm now down to one GTX295 per System, as this was the only config I could get to work after over a week of intense troubleshooting (two Systems run XP32, one System runs Vista64)
Seeing both Vista64 and Win7 64 completely fail to have the Win64 185.85WHQL Driver report more than 2 GPUs to BOINC for CUDA use, I'm almost confident it's a Driver issue.
The same occured when I enabled the CUDA capable IGP; point blank fresh install with all needed Service Pack and System Drivers resulted a fresh BOINC 6.6.20 (32bit and 64bit) Install to see only the IGP, despite upto 2 GTX295 being active and seen by the Driver panel.
I also found some serious SLI performance problems with the Win64 185.85WHQL Driver under DX9 and especially OpenGL, all cards were working fine (at least the hardware), supporting my suspicion that there's something buggy buried in that Driver.
The 32bit 185.85WHQL Driver apparently does not have any such limitation, under WinXP32 immediately reported all 4 GPUs to BOINC as it should (useless with <400MB usable RAM remaining, but technically it did not have any problems).
-
Still no luck with the 295, because I have the same problem.
-
There are 2 potential fixes for this and it's not just Windows 7 but x64 Vista as well.
There is a new command on BOINC in cc_config.xml to avoid the incorrect memory sizing on the second GPU or you can add LimitVideoPresentSources to the registry for th card. This stops the need for any dummy VGA adapter etc. The problem I had is due to Vista and Windows 7 only enabling the GPU if there is a device connected to the VGA/DVI port.
The xml to add to cc_config.xml is:-
<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
</options>
</cc_config>
cc_config.xml (http://boinc.berkeley.edu/wiki/Client_configuration) file definition is here.
Full solution for LimitVideoPresentSources is at the BOINC Forums (http://boinc.berkeley.edu/dev/forum_thread.php?id=3825#25150)
-
There are 2 potential fixes for this and it's not just Windows 7 but x64 Vista as well.
There is a new command on BOINC in cc_config.xml to avoid the incorrect memory sizing on the second GPU or you can add LimitVideoPresentSources to the registry for th card. This stops the need for any dummy VGA adapter etc. The problem I had is due to Vista and Windows 7 only enabling the GPU if there is a device connected to the VGA/DVI port.
The xml to add to cc_config.xml is:-
<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
</options>
</cc_config>
cc_config.xml (http://boinc.berkeley.edu/wiki/Client_configuration) file definition is here.
Full solution for LimitVideoPresentSources is at the BOINC Forums (http://boinc.berkeley.edu/dev/forum_thread.php?id=3825#25150)
I did all that..(<use_all_gpus>1</use_all_gpus>, plugged in a monitor to the 2e screen, extended the desktop, even tried the registry) but no luck. The card works fine in XP.
-
Also doesn't work for me - BOINC 6.6.36, NVidia 186.18 with 2 GTX280s...Only sees one...Worked fine under Vista 64 bit...
-
If BOINC sees only 1 GPU report this on BOINC site (questions and problems forum).
Also it's worth to search nVidia's forums for known problems with Windows 7.
-
If anyone stumbles on this post, the latest beta 190.38 NVidia drivers fix this issue -
http://www.nvidia.com/Download/Find.aspx?lang=en-us
-
If anyone stumbles on this post, the latest beta 190.38 NVidia drivers fix this issue -
http://www.nvidia.com/Download/Find.aspx?lang=en-us
Yes this seems to work. I now See 2 cards on my 295.
But I still had to plug in a second screen.
What I only lack is work...
-
I only have one display plugged in with 2 280s - 295 may be a bit different...
-
I only have one display plugged in with 2 280s - 295 may be a bit different...
It seems to work with one as well. But Win 7 been down for some weeks so he request only a max of 10 WU. ;D Not much to go on.
And of the 10 I got only 2 completed without an error. So that's not very promising.
-
I only have one display plugged in with 2 280s - 295 may be a bit different...
It seems to work with one as well. But Win 7 been down for some weeks so he request only a max of 10 WU. ;D Not much to go on.
And of the 10 I got only 2 completed without an error. So that's not very promising.
This doesn't work with all 295 card, maybe only with certain revisions of the board.