can I have it both ways? (fast and slow GPU)

Message boards : Number crunching : can I have it both ways? (fast and slow GPU)
Message board moderation

To post messages, you must log in.

AuthorMessage
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 389
Credit: 236,772,605
RAC: 374
United States
Message 1143154 - Posted: 20 Aug 2011, 19:21:52 UTC

I Have a "fast" and a "slow" GPU running MB under Lunatic 0.38 setup 64 Bit Boinc 6.12.33.
They are:

NVIDIA GPU 0: GeForce GTX 550 Ti (driver version 27533, CUDA version 4000, compute capability 2.1, 961MB, 461 GFLOPS peak)
NVIDIA GPU 1: GeForce 9500 GT (driver version 27533, CUDA version 4000, compute capability 1.1, 967MB, 90 GFLOPS peak)

There is about 33% of idle time, (+/-), on the GTX 550 Ti but none on the 9500 GT.

When I put <count>.5</count> in the 9500 also gets 2 WU's

Is there a way to just give the GTX 550 2 WU's and just 1 to the 9500??

Thanks, again

Ed F

ID: 1143154 · Report as offensive
Cruncher-American Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor

Send message
Joined: 25 Mar 02
Posts: 1513
Credit: 370,893,186
RAC: 340
United States
Message 1143167 - Posted: 20 Aug 2011, 19:59:11 UTC

I do not think you can do that. But you can only use the faster card and ignore the slower card which really is a helluva lot slower. Then you can use the slower card for normal graphics and let the gtx go crazy with SETI.
ID: 1143167 · Report as offensive
Profile skildude
Avatar

Send message
Joined: 4 Oct 00
Posts: 9541
Credit: 50,759,529
RAC: 60
Yemen
Message 1143169 - Posted: 20 Aug 2011, 20:05:48 UTC - in response to Message 1143167.  

if you install the lunatics app you'll be able to run 2-3 WU's on a 5XX GPU which actually defeats having the second one onboard


In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope
ID: 1143169 · Report as offensive
Claggy
Volunteer tester

Send message
Joined: 5 Jul 99
Posts: 4654
Credit: 47,537,079
RAC: 4
United Kingdom
Message 1143174 - Posted: 20 Aug 2011, 20:18:43 UTC - in response to Message 1143154.  

Run a second Boinc client with just the slow GPU enabled in it, see Vyper's Blog on how to do it:

Running Nvidia Different architectures optimal in same host with SETI @ Home and Other BOINC projects

Claggy
ID: 1143174 · Report as offensive
Cruncher-American Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor

Send message
Joined: 25 Mar 02
Posts: 1513
Credit: 370,893,186
RAC: 340
United States
Message 1143213 - Posted: 20 Aug 2011, 22:19:56 UTC - in response to Message 1143174.  

Run a second Boinc client with just the slow GPU enabled in it, see Vyper's Blog on how to do it:

Running Nvidia Different architectures optimal in same host with SETI @ Home and Other BOINC projects

Claggy



You CAN do this, of course, but is it really worth it, given the difference in capabilities of the two cards?
ID: 1143213 · Report as offensive
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 389
Credit: 236,772,605
RAC: 374
United States
Message 1143262 - Posted: 21 Aug 2011, 1:06:28 UTC - in response to Message 1143213.  
Last modified: 21 Aug 2011, 1:08:12 UTC

It makes a wonderful thought experiment ... but what I was wanting to know (if it exists) is how to run 2 WU's on the fast card and 1 on the slow ... running 2 on the slow card works ... it just seems ... well ... silly I guess. The running software calls the 9500GT a cuda fermi ( witch it isn't) ...

There just seems to be soooooo many neat little < ...> x </ ...> commands that can be included on the setup that perhaps there was one for this situation.

(BTW where can I get documentation on the config (et al) qualifiers?

Ed F

p.s. every little WU done for the team is one more closer to the goal.
ID: 1143262 · Report as offensive
Profile BilBg
Volunteer tester
Avatar

Send message
Joined: 27 May 07
Posts: 3720
Credit: 9,385,827
RAC: 0
Bulgaria
Message 1143268 - Posted: 21 Aug 2011, 1:20:20 UTC - in response to Message 1143262.  
Last modified: 21 Aug 2011, 1:26:25 UTC


"... how to run 2 WU's on the fast card and 1 on the slow" - the link gives explanation for exactly this (by running 2 BOINC clients at the same time).

"The purpose of this was in my case to have the orginal installation modifed to run the Fermi board and the CPU’s most optimal and the other installation would run my Nvidia 2xx generation card by itself..

In that way i could configure app_info.xml to make the fermi board run two WU’s at once and not even bother trying to use the 2xx board for crunching and 2xx board 1 WU at once with another executable which is more suitable for that board."


You can't do that with any other config/option/preference.


"BTW where can I get documentation on the config (et al) qualifiers?" - if you don't know that the "two BOINC clients at the same time" is maybe not for you - it involves some advanced actions.

http://boinc.berkeley.edu/wiki/Client_configuration
http://boincfaq.mundayweb.com/index.php?language=1&view=91
http://www.boinc-wiki.info/Cc_config.xml


 


- ALF - "Find out what you don't do well ..... then don't do it!" :)
 
ID: 1143268 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1143344 - Posted: 21 Aug 2011, 6:34:19 UTC - in response to Message 1143262.  
Last modified: 21 Aug 2011, 6:38:04 UTC

It makes a wonderful thought experiment ... but what I was wanting to know (if it exists) is how to run 2 WU's on the fast card and 1 on the slow ... running 2 on the slow card works ... it just seems ... well ... silly I guess. The running software calls the 9500GT a cuda fermi ( witch it isn't) ...

There just seems to be soooooo many neat little < ...> x </ ...> commands that can be included on the setup that perhaps there was one for this situation.

(BTW where can I get documentation on the config (et al) qualifiers?

Ed F

p.s. every little WU done for the team is one more closer to the goal.


You are not the first who would like to crunch with non- and Fermi cards in one machine.. ;-)

So we know from what we speak.. ;-)

2 CUDA WUs/non-Fermi GPU will slow down a lot.

To now, the BOINC DEVs didn't included a feature for to let run 2+ WUs/Fermi GPU and 1 WU/non-Fermi GPU. And it looks like this will happen never.
So you would need to use a work around.

So the only way would be to let run two BOINCs simultaneously (like already mentioned here in this thread).

Normally the 9500GT is disabled automatically then in one BOINC (because if the GPUs are to different the fastest is used). There the CPU and the GTX550Ti calculate WUs.
In the other BOINC you would need to disable the GTX550Ti via cc_config.xml file. There then only the 9500GT calculate WUs.
(Over project settings and/or cc_config.xml file.)


- Best regards! - Sutaru Tsureku, team seti.international founder. - Optimize your PC for higher RAC. - SETI@home needs your help. -
ID: 1143344 · Report as offensive
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 389
Credit: 236,772,605
RAC: 374
United States
Message 1143677 - Posted: 22 Aug 2011, 2:34:21 UTC - in response to Message 1143344.  

Thanks to all!

I'm not detoured by the 2 boinc's ... it sounds like fun ... but is really worth the effort beyond the educational value ( .... maybe yes ... maybe no ... I'll give it some thought).

It's always fun to discover the ins and outs of an application ... one of this complexity is large enough to let one marvel at the work done for OUR benefit!

I'm not ready to master ANYTHING in BOINC/SETI but it is fun to gaze and wonder!!!

Thanks to all who have mastered the joys of discovery here ... thanks for your help and patience!

Ed F
ID: 1143677 · Report as offensive
Profile skildude
Avatar

Send message
Joined: 4 Oct 00
Posts: 9541
Credit: 50,759,529
RAC: 60
Yemen
Message 1143684 - Posted: 22 Aug 2011, 2:40:49 UTC - in response to Message 1143677.  

honestly though an nVidia 8XXX is a pig compared to the 5XX cards. if it were a 2XX series card it might be worth your while. as it stands the 8XXX card is more of a waste of electricity that just setting up the 5XX and letting it rip doing 2 WU's at a time


In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope
ID: 1143684 · Report as offensive
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 389
Credit: 236,772,605
RAC: 374
United States
Message 1143700 - Posted: 22 Aug 2011, 3:09:31 UTC - in response to Message 1143684.  

'ya I think this makes good sense ...

I think for the time being, I'll set up the fast card for 2WU's and the slow card to be my video card ... I don't do games more demanding then "free cell" anyway.

Once I get a feeling for set up, then I can play a little.

Ed F
ID: 1143700 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1143738 - Posted: 22 Aug 2011, 5:32:03 UTC - in response to Message 1143700.  

Don't be pessimistic.. ;-)

IIRC, a 9800 is fine for the half of a GTX260, so maybe ~ 6,000 RAC.
Then your 9500GT maybe ~ 2,000 to 4,000 RAC. It's more than other members have in a whole.

http://www.nvidia.com/object/product_geforce_9500gt_us.html
32 CUDA cores and max 50 W, so if CUDA then maybe ~ 38 W (~ 3/4 of the max from my experiences).
Not so much wasted electricity.. ;-)

Try it and tell us your experiences with two BOINCs simultaneously.. ;-)


- Best regards! - Sutaru Tsureku, team seti.international founder. - Optimize your PC for higher RAC. - SETI@home needs your help. -
ID: 1143738 · Report as offensive
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 389
Credit: 236,772,605
RAC: 374
United States
Message 1144563 - Posted: 24 Aug 2011, 14:28:16 UTC
Last modified: 24 Aug 2011, 14:29:27 UTC

O.k. I've Been running 1 fast and 1 slow GPU now for a while and I see the "remaining" time est adjust up and down as Boinc attempts to est the avg speed so the correct number of WU's can be downloaded etc.

However, why does this ALSo affect the CPU scheduling??

Is there a separate "<duration_correction_factor>" for the GPU's??

My <duration_correction_factor> has gone from 1.0 to 4.864084 since I turned the "slow" GPU on, and so am out of CPU WU' but have some GPU WU's left (SOB!!) Is this an intended affect or an oversight when they implemented GPU's??

Ed F
ID: 1144563 · Report as offensive
Josef W. Segur
Volunteer developer
Volunteer tester

Send message
Joined: 30 Oct 99
Posts: 4504
Credit: 1,414,761
RAC: 0
United States
Message 1144625 - Posted: 24 Aug 2011, 17:37:55 UTC - in response to Message 1144563.  

O.k. I've Been running 1 fast and 1 slow GPU now for a while and I see the "remaining" time est adjust up and down as Boinc attempts to est the avg speed so the correct number of WU's can be downloaded etc.

However, why does this ALSo affect the CPU scheduling??

Is there a separate "<duration_correction_factor>" for the GPU's??

My <duration_correction_factor> has gone from 1.0 to 4.864084 since I turned the "slow" GPU on, and so am out of CPU WU' but have some GPU WU's left (SOB!!) Is this an intended affect or an oversight when they implemented GPU's??

Ed F

No, there's just one client side <duration_correction_factor> (DCF) per project. The issue wasn't ignored, see http://boinc.berkeley.edu/trac/wiki/RuntimeEstimation for the adaptation Dr. Anderson implemented. However, also note that the line which says "DCF is no longer used." really means the server-side software ignores it although the client still uses and adjusts it.

In a stable situation the server adjustments do tend to force DCF to about 1.0 particularly for those running stock, and well enough for those running anonymous platform. Transients like adding a slower GPU cause an excursion which takes over a hundred validated tasks from the new configuration for adaptation.

It is difficult to get right, but using <flops> in the app_info.xml does provide the potential to avoid such excursions. That's not a recommendation, doing so requires thinking ahead and making <flops> adjustments at the same time hardware changes or speed tweaks are applied.

Under the present cicumstances, even correcting DCF probably won't get any CPU work for your hosts. If your host asks for both CPU and GPU work the servers will first assign any tasks which are available to the GPUs, and it is unlikely there will be enough to even satisfy the GPU part of the request.
                                                                  Joe
ID: 1144625 · Report as offensive
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 389
Credit: 236,772,605
RAC: 374
United States
Message 1144632 - Posted: 24 Aug 2011, 17:56:26 UTC - in response to Message 1144625.  

Thanks to the "team" again. I appreciate the time you put in answering our questions!

Thanks, again.

Ed F
ID: 1144632 · Report as offensive
EdwardPF
Volunteer tester

Send message
Joined: 26 Jul 99
Posts: 389
Credit: 236,772,605
RAC: 374
United States
Message 1147349 - Posted: 31 Aug 2011, 3:46:46 UTC - in response to Message 1144632.  

Just another observation,

After running with both GPU's the 9500 and the 550Ti I got 11 error -177 (timeout) on the 9500 GPU. it seems the diff between them didn't yield a usable APR(?)

Ed F
ID: 1147349 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 1147469 - Posted: 31 Aug 2011, 13:17:56 UTC - in response to Message 1147349.  
Last modified: 31 Aug 2011, 13:33:31 UTC

This -177 error happens if the computation time reached a value much more than the estimate time was.

For to prevent this, you should use Fred's BoincRescheduler.
This tool prevent the -177 error.

I guess you would need to do this only in the BOINC folder of your 9500GT.

But, you shouldn't longer use this tool for to reschedule WUs from CPU to GPU or the other way, because this would disturb now the calculation of the Cr./WU (CreditNew).


[EDIT: Also you should make a config.xml file with:
<config>
   <seti>  
      <vlar>0.12</vlar>
   </seti>
</config>

Because the default value in this tool is an other as S@h use as nVIDIA GPU VLAR value. So the tool would say 'rescheduling needed' althought it's not needed.]


- Best regards! - Sutaru Tsureku, team seti.international founder. - Optimize your PC for higher RAC. - SETI@home needs your help. -
ID: 1147469 · Report as offensive

Message boards : Number crunching : can I have it both ways? (fast and slow GPU)


 
©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.