3WU BRP3cuda on a single GPU

dunx
dunx
Joined: 13 Aug 10
Posts: 119
Credit: 27,676,692
RAC: 0

This is beyond my skill

This is beyond my skill level, but...

You could install another BOINC "set-up" for the old-old card and run an APP on it.

The normal BOINC set-up just needs telling to ignore the n-th GPU...

Hope this makes sense. I read it somewhere but lost the link... us old folk eh ?

dunx

Jeroen
Jeroen
Joined: 25 Nov 05
Posts: 379
Credit: 738,419,435
RAC: 0

RE: notably with 8 HT

Quote:


notably with 8 HT cores still doing 8 non-CUDA tasks as well! So now I have to cook up an app_info ??? Probably can get 2 + 1 + 2 = 5 simultaneous CUDA's at least ..... ;-)

Cheers, Mike.

You may be able to reduce your GPU processing time some by scaling back on the CPU work units. This is specifically because of some performance loss seen by HT and the GPU application having dependence on the CPU. I keep my i7 CPU at less than or equal to 50% load for this reason. I have seen some decent gains in GPU processing performance by doing this.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,024
Credit: 104,149,380
RAC: 90,259

Thank you Richard, dunx and

Thank you Richard, dunx and Jeroen ! :-)

So I could :

- use the old_old 9500 to run my display alone, but I have to ....

- rearrange the cards in the slots ( darn windows insists on the primary/only display being the one in the first slot, plus I don't have onboard video ), but ......

- run down the work first so as to not stuff up my wing-people. I know Gary has mentioned the workaround for that but I was going to be more incremental by first watching temps especially, plus while my Corsair 750W PSU is quite good but let's just edge him up instead, then ....

- tell BOINC to ignore the old_old 512 MB card ( how do I write that into the cc_config ? ) which will now be indexed as the zero'th GPU, then ...

- app_info.xml my way to allowing the other two 1G cards to accept work progressing up to three CUDA's each ?

- scale back on CPU-only units to allows the GPU units better service/throughput ( more diners in the restaurant means more waiters )

[ AIM : grand total of 2 x 3 CUDA's simultaneously. I guess users with video-on-mobo and 3 PCI-E slots, plus 3 x 1G NVidia cards of course, could go for a niner .... hey, a mate of mine has got a ..... :-) :-) ]

BTW : If anyone has Windows 7 Ultimate I've noted :

(a) If you come into it via Remote Desktop then BOINC ( acting on Windows advice ? ) marks the GPU's as 'unusable' for the session ( and noted thus in the message log and in the task pane ), and GPU processing is paused until you log out. No problems seem to be generated by this for me, but I found that of interest. I was checking my progress from elsewhere and noticed that with some initial distress.

(b) The BOINC data directory doesn't show in the default viewing of Explorer ie. if you don't go to 'Folder Options' in Control Panel and select 'Show hidden files, folders, and drives'. If it's hidden then Windows search facility won't search it either. [ There's fifteen minutes I'll never get back, I want the old Windoze search facility I was so used to. ] For my machine you'd be after 'C:\ProgramData\BOINC' to put 'cc_config.xml' into and 'C:\ProgramData\BOINC\projects\einstein.phys.uwm.edu' for 'app_info.xml'.

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter. Blaise Pascal

dunx
dunx
Joined: 13 Aug 10
Posts: 119
Credit: 27,676,692
RAC: 0

Hello again ! There is

Hello again !

There is also a "hit" depending on your MoBo's PCI-E lane distribution to said graphics cards.

I'm sorry but I'm not sure what you are using, BUT on my Asus Maximus III the available PCI-E lanes were split from 1 x 16, to 2x 8, to x8,x8,x4.

GTX 460 OC 1 WU = 46 minutes
GTX 460 OC 2 WU = 66 minutes
GTX 460 OC x2 @ 2 WU = 120 minutes - YES that's half as fast ! ! !

As you can see this halved the work-rate ! ! ! So I put one card in each PC and doubled the work done per day. And started "digging"...

Now running an Asus P6T6 WS that has 3 of 16X PCI-E graphics slots, now doing four WU in 66 minutes, 2 per GTX 460... now need a third, LoL.

So adding GPU's doesn't always pay off as expected...

Plus as Jeroen stated, I run 6 CPU threads on MW@Home and leave 2 threads "free" for surfing the web and to feed the two GTX 460's.

HTH

dunx

P.S. I find that W7 and the hardware don't always agree as to which GPU is which...[/b]

Gundolf Jahn
Gundolf Jahn
Joined: 1 Mar 05
Posts: 1,079
Credit: 341,280
RAC: 0

RE: - tell BOINC to ignore

Quote:
- tell BOINC to ignore the old_old 512 MB card ( how do I write that into the cc_config ? ) which will now be indexed as the zero'th GPU, then ...


That would be[pre]0[/pre]as of Client configuration in the BOINC wiki. But, as dunx mentioned, check that BOINC ignores the right one. :-)

Gruß,
Gundolf

Computer sind nicht alles im Leben. (Kleiner Scherz)

dunx
dunx
Joined: 13 Aug 10
Posts: 119
Credit: 27,676,692
RAC: 0
Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,024
Credit: 104,149,380
RAC: 90,259

OK, three cards doesn't work.

OK, three cards doesn't work. I get a bunch of errors like :
[pre]6.10.60

Cannot create a symbolic link in a registry key that already has subkeys or values. (0x3fc) - exit code 1020 (0x3fc)

Activated exception handling...
[00:13:31][5332][INFO ] Starting data processing...
[00:13:31][5332][ERROR] Couldn't initialize CUDA driver API (error: 100)!
[00:13:31][5332][ERROR] Demodulation failed (error: 1020)!
00:13:31 (5332): called boinc_finish

]]>[/pre]
So I take the old old card out. Much the same happened with two cards. So now I'm using only one of the two remaining in the machine. The one that's running CUDA WU's ( NVIDIA GPU 1 ) is not the one running the display ( NVIDIA GPU 0 ), as I cc_config'ed that out of play ( thank you Gundolf ) :

[pre]

0

[/pre]
[pre]26/04/2011 2:06:11 PM NVIDIA GPU 0 (ignored by config): GeForce GTX 560 Ti (driver version 26658, CUDA version 3020, compute capability 2.1, 993MB, 901 GFLOPS peak)
26/04/2011 2:06:11 PM NVIDIA GPU 1: GeForce 9800 GT (driver version 26658, CUDA version 3020, compute capability 1.1, 998MB, 336 GFLOPS peak)[/pre]
and so far so good .... until, being a suspicious fellow, I decided to Remote Desktop into this Win7 Ultimate machine from elsewhere in the house. Then this and that happened. Here's the relevant section of the BOINC message pane for that period :

[pre]26/04/2011 5:47:18 PM GPUs have become unusable; disabling tasks
26/04/2011 5:47:41 PM GPUs have become usable; enabling tasks
26/04/2011 5:47:42 PM Einstein@Home Restarting task PM0091_00531.dm_132_0 using einsteinbinary_BRP3 version 107
26/04/2011 5:47:45 PM Einstein@Home Computation for task PM0091_00531.dm_132_0 finished
26/04/2011 5:47:45 PM Einstein@Home Output file PM0091_00531.dm_132_0_1 for task PM0091_00531.dm_132_0 absent
26/04/2011 5:47:45 PM Einstein@Home Output file PM0091_00531.dm_132_0_2 for task PM0091_00531.dm_132_0 absent
26/04/2011 5:47:45 PM Einstein@Home Output file PM0091_00531.dm_132_0_3 for task PM0091_00531.dm_132_0 absent
26/04/2011 5:47:45 PM Einstein@Home Starting PM0091_00531.dm_404_0
26/04/2011 5:47:46 PM Einstein@Home Starting task PM0091_00531.dm_404_0 using einsteinbinary_BRP3 version 107
26/04/2011 5:47:47 PM Einstein@Home Started upload of PM0091_00531.dm_132_0_0
26/04/2011 5:47:49 PM Einstein@Home Computation for task PM0091_00531.dm_404_0 finished
26/04/2011 5:47:49 PM Einstein@Home Output file PM0091_00531.dm_404_0_0 for task PM0091_00531.dm_404_0 absent
26/04/2011 5:47:49 PM Einstein@Home Output file PM0091_00531.dm_404_0_1 for task PM0091_00531.dm_404_0 absent
26/04/2011 5:47:49 PM Einstein@Home Output file PM0091_00531.dm_404_0_2 for task PM0091_00531.dm_404_0 absent
26/04/2011 5:47:49 PM Einstein@Home Output file PM0091_00531.dm_404_0_3 for task PM0091_00531.dm_404_0 absent
26/04/2011 5:47:49 PM Einstein@Home Starting PM0092_01231.dm_272_0
26/04/2011 5:47:50 PM Einstein@Home Starting task PM0092_01231.dm_272_0 using einsteinbinary_BRP3 version 107
26/04/2011 5:47:51 PM Einstein@Home Finished upload of PM0091_00531.dm_132_0_0
26/04/2011 5:48:51 PM Einstein@Home Sending scheduler request: To fetch work.
26/04/2011 5:48:51 PM Einstein@Home Reporting 2 completed tasks, requesting new tasks for GPU[/pre]
Sooooo : two cards in the machine, only using one card, only one task at a time on that card plus it's not even handling my display, plenty of memory available on the card, throw in a remote desktop call, the running task dm_132 falls over, then dm_404 after that folds up after a mere few seconds ??? My problem is not evidently multiple GPU's or multiple tasks per GPU. Some sort of unrecoverable context interruption triggered by the Remote Desktop session. So I revise my earlier comment "No problems seem to be generated by this" .... :-)

Cheers, Mike.

( edit) Have I tripped up(on) a corner case ??

( edit ) Or "I find that W7 and the hardware don't always agree as to which GPU is which" ( thank you Dunx ) is perhaps a fairly floating proposition ? That is : W7 won't restore a CUDA context to a non-displaying card ?

I have made this letter longer than usual because I lack the time to make it shorter. Blaise Pascal

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 1,914
Credit: 133,021,074
RAC: 109,841

RE: Sooooo : two cards in

Quote:


Sooooo : two cards in the machine, only using one card, only one task at a time on that card plus it's not even handling my display, plenty of memory available on the card, throw in a remote desktop call, the running task dm_132 falls over, then dm_404 after that folds up after a mere few seconds ??? My problem is not evidently multiple GPU's or multiple tasks per GPU. Some sort of unrecoverable context interruption triggered by the Remote Desktop session. So I revise my earlier comment "No problems seem to be generated by this" .... :-)

Cheers, Mike.

( edit) Have I tripped up(on) a corner case ??

( edit ) Or "I find that W7 and the hardware don't always agree as to which GPU is which" ( thank you Dunx ) is perhaps a fairly floating proposition ? That is : W7 won't restore a CUDA context to a non-displaying card ?


You've found another well-known Gotcha for first-time CUDA users. Remote Desktop is incompatible with CUDA crunching - RDP unloads the nVidia drivers, and replaces them with Microsoft drivers which don't support CUDA. (That's on Windows 7 and Vista, with the new secure driver model - XP is OK). It's somewhat related to the reason you can't run CUDA in BOINC's 'service' or 'Protected Application Execution' mode on Windows 7 and Vista.

Other remote access products such as the various flavours of VNC are known to work - we ought to draw up a list sometime.

Edit - and to complete the set of incompatibilities before you fall over the next one - Windows fast user switching tends to kill CUDA tasks too, although the BOINC developers are trying to make it suspend/resume the task, rather than have it error out.

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6,024
Credit: 104,149,380
RAC: 90,259

RE: You've found another

Quote:

You've found another well-known Gotcha for first-time CUDA users. Remote Desktop is incompatible with CUDA crunching - RDP unloads the nVidia drivers, and replaces them with Microsoft drivers which don't support CUDA. (That's on Windows 7 and Vista, with the new secure driver model - XP is OK). It's somewhat related to the reason you can't run CUDA in BOINC's 'service' or 'Protected Application Execution' mode on Windows 7 and Vista.

Other remote access products such as the various flavours of VNC are known to work - we ought to draw up a list sometime.

Edit - and to complete the set of incompatibilities before you fall over the next one - Windows fast user switching tends to kill CUDA tasks too, although the BOINC developers are trying to make it suspend/resume the task, rather than have it error out.

Thank you Richard! I've re-invented the square wheel. Oh well ... :-)

Cheers, Mike.

I have made this letter longer than usual because I lack the time to make it shorter. Blaise Pascal

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.