Running multiple GPU tasks on one video card.


log in

Advanced search

Message boards : Number crunching : Running multiple GPU tasks on one video card.

1 · 2 · 3 · Next
Author Message
spitfire_mk_2
Avatar
Send message
Joined: 14 Apr 00
Posts: 458
Credit: 12,797,928
RAC: 8,803
United States
Message 1234189 - Posted: 20 May 2012, 20:43:48 UTC

I do seti@home and Einstein@home.

Einstein@home has this neat thing in the project preferences where you can tell it how many instances of cuda task you want your video card to do.

seti@home does not have this feature at this time. Is there a howto somewhere that tells how to setup so that my video card does more than one cuda work unit?

I have looked at Lunatics download site. They got stuff, but things are a bit convoluted in there. I am looking at something cleaner.
____________

Profile Area 51
Avatar
Send message
Joined: 31 Jan 04
Posts: 965
Credit: 42,193,520
RAC: 0
United Kingdom
Message 1234194 - Posted: 20 May 2012, 20:53:04 UTC - in response to Message 1234189.
Last modified: 20 May 2012, 20:54:08 UTC

I do seti@home and Einstein@home.

Einstein@home has this neat thing in the project preferences where you can tell it how many instances of cuda task you want your video card to do.

seti@home does not have this feature at this time. Is there a howto somewhere that tells how to setup so that my video card does more than one cuda work unit?

I have looked at Lunatics download site. They got stuff, but things are a bit convoluted in there. I am looking at something cleaner.



you will need to edit your app_info.xml. Replace all instances of <count>1</count> with the reciprocal of how many tasks you want to run per GPU (ie, 0.5 for two tasks, 0.3 for 3 etc etc). Don't forget to stop your host processing whilst you do this.
____________

spitfire_mk_2
Avatar
Send message
Joined: 14 Apr 00
Posts: 458
Credit: 12,797,928
RAC: 8,803
United States
Message 1234203 - Posted: 20 May 2012, 21:07:02 UTC - in response to Message 1234194.

I do seti@home and Einstein@home.

Einstein@home has this neat thing in the project preferences where you can tell it how many instances of cuda task you want your video card to do.

seti@home does not have this feature at this time. Is there a howto somewhere that tells how to setup so that my video card does more than one cuda work unit?

I have looked at Lunatics download site. They got stuff, but things are a bit convoluted in there. I am looking at something cleaner.



you will need to edit your app_info.xml. Replace all instances of <count>1</count> with the reciprocal of how many tasks you want to run per GPU (ie, 0.5 for two tasks, 0.3 for 3 etc etc). Don't forget to stop your host processing whilst you do this.

I don't have app_info.xml. I am running unaltered boinc 7.0.25. That is why I am looking for a good howto.

____________

Profile Area 51
Avatar
Send message
Joined: 31 Jan 04
Posts: 965
Credit: 42,193,520
RAC: 0
United Kingdom
Message 1234208 - Posted: 20 May 2012, 21:17:46 UTC - in response to Message 1234203.

I do seti@home and Einstein@home.

Einstein@home has this neat thing in the project preferences where you can tell it how many instances of cuda task you want your video card to do.

seti@home does not have this feature at this time. Is there a howto somewhere that tells how to setup so that my video card does more than one cuda work unit?

I have looked at Lunatics download site. They got stuff, but things are a bit convoluted in there. I am looking at something cleaner.



you will need to edit your app_info.xml. Replace all instances of <count>1</count> with the reciprocal of how many tasks you want to run per GPU (ie, 0.5 for two tasks, 0.3 for 3 etc etc). Don't forget to stop your host processing whilst you do this.

I don't have app_info.xml. I am running unaltered boinc 7.0.25. That is why I am looking for a good howto.



Since your computers are hidden, and that was not mentioned.....

The more information you give, the more likely you are to get an appropriate response.
____________

Profile Fred E.Project donor
Volunteer tester
Send message
Joined: 22 Jul 99
Posts: 768
Credit: 24,139,004
RAC: 1
United States
Message 1234213 - Posted: 20 May 2012, 21:22:43 UTC - in response to Message 1234203.

I do seti@home and Einstein@home.

Einstein@home has this neat thing in the project preferences where you can tell it how many instances of cuda task you want your video card to do.

seti@home does not have this feature at this time. Is there a howto somewhere that tells how to setup so that my video card does more than one cuda work unit?

I have looked at Lunatics download site. They got stuff, but things are a bit convoluted in there. I am looking at something cleaner.



you will need to edit your app_info.xml. Replace all instances of <count>1</count> with the reciprocal of how many tasks you want to run per GPU (ie, 0.5 for two tasks, 0.3 for 3 etc etc). Don't forget to stop your host processing whilst you do this.

I don't have app_info.xml. I am running unaltered boinc 7.0.25. That is why I am looking for a good howto.



The anonymous platform is the only method available here. Einstein's method uses modified applications and special tasks. They offered their method for adoption by other BOINC projects, but Dr. Anderson did not approve of their methodology.

The Lunatics' apps don't modify BOINC - they replace the stock applications provided by the project. The installer will help you select the right application to get optimum results from your cpu, etc. Can't get specific in your case because your computers are hidden and I can't see your platform. They don't impact other projects. I'm currently running 3 at a time at SETI using app_info.xml and 3 at a time at Einstein using their method.
____________
Another Fred
Support SETI@home when you search the Web with GoodSearch or shop online with GoodShop.

spitfire_mk_2
Avatar
Send message
Joined: 14 Apr 00
Posts: 458
Credit: 12,797,928
RAC: 8,803
United States
Message 1234219 - Posted: 20 May 2012, 21:34:38 UTC
Last modified: 20 May 2012, 21:36:31 UTC

Everyone, thank you for taking your time.

What I have:
intel Core 2 Quad
Windows Xp 32-bit
nVidia GTX 460
BOINC 7.0.25

I am guessing I need this auto installer: http://lunatics.kwsn.net/index.php?module=Downloads;sa=dlview;id=329
But the Readme says they are still working on version for BOINC 7, so I am not sure.
____________

Profile Fred E.Project donor
Volunteer tester
Send message
Joined: 22 Jul 99
Posts: 768
Credit: 24,139,004
RAC: 1
United States
Message 1234223 - Posted: 20 May 2012, 21:52:54 UTC

Looks like the right one - version 0.40 for 32 bit,

The reference to development for v7 is for the upcoming SETI multibeam v7 application, not the same as BOINC v7.xx that you are running. As the readme says, we'll have to update when the project has the new apps ready.
____________
Another Fred
Support SETI@home when you search the Web with GoodSearch or shop online with GoodShop.

spitfire_mk_2
Avatar
Send message
Joined: 14 Apr 00
Posts: 458
Credit: 12,797,928
RAC: 8,803
United States
Message 1234232 - Posted: 20 May 2012, 22:19:16 UTC

msattler, Fred E. thank you very much, it is up and running.
____________

AlProject donor
Send message
Joined: 3 Apr 99
Posts: 481
Credit: 53,426,045
RAC: 29,084
United States
Message 1234350 - Posted: 21 May 2012, 2:36:45 UTC

Not to jack the thread, but is there a good rule of thumb to know how many tasks a given card can (or should comfortably be able to) process at one time? I'm all for using hardware to it's fullest and most productive purpose, especially when it has recently been upgraded. ;-)
____________

Profile Rick
Send message
Joined: 27 Mar 01
Posts: 8
Credit: 21,615,319
RAC: 0
United States
Message 1234374 - Posted: 21 May 2012, 3:43:06 UTC

@Al - It's generally regarded that one instance per 512MB of VRAM (provided its not a 3GB GT 430 which exists, I dont think six instances here would work out so well) is an acceptable comfort zone. Bottom line, if you can run more without causing 2d lag, you should be fine.

jravin
Send message
Joined: 25 Mar 02
Posts: 941
Credit: 102,230,270
RAC: 85,470
United States
Message 1234381 - Posted: 21 May 2012, 3:55:24 UTC - in response to Message 1234350.

Not to jack the thread, but is there a good rule of thumb to know how many tasks a given card can (or should comfortably be able to) process at one time? I'm all for using hardware to it's fullest and most productive purpose, especially when it has recently been upgraded. ;-)


One thing you can do is get some freeware that tells you GPU utilization - NVIDIA has one, as do other manufacturers - and, if it is not near 100%, increase (by 1) the number of WUs run at a time. Iterate until utilization is near 100%. (Once ut'n is near 100%, you can't get more work done by increasing the WUs).

But keep in mind what Rick said - you don't want to run out of GPU memory either.

____________

fataldog187
Send message
Joined: 4 Nov 02
Posts: 42
Credit: 1,271,261
RAC: 0
United States
Message 1234401 - Posted: 21 May 2012, 5:33:29 UTC - in response to Message 1234381.

Once ut'n is near 100%, you can't get more work done by increasing the WUs).

Good to know, I was thinking of trying 2 work units on my GPU since it has 1GB VRAM but it's running at 96-99% all the time so I guess I'll hold off til I get a more powerful card for that.

You would think there would be a thread with basic guidelines from other people listing their GPU's and what works for them so that there isn't as much trial and error involved.
____________

AlProject donor
Send message
Joined: 3 Apr 99
Posts: 481
Credit: 53,426,045
RAC: 29,084
United States
Message 1234475 - Posted: 21 May 2012, 12:11:39 UTC - in response to Message 1234374.

Thanks for the reply. I found the file, but have another question. I know my new 560's have 1.2 gig of ram, not sure about the 285, but I know the 260 doesn't have a ton of ram, so I would like to start out by having only the 2 560's doing multiple units. I read thru the file as best as I could but it appeared that the <count>1</count> settings were per app, not per card. Is it possible to spec that setting on each piece of hardware, or is it a global setting per application? Thanks!
____________

tbretProject donor
Volunteer tester
Avatar
Send message
Joined: 28 May 99
Posts: 2859
Credit: 214,881,922
RAC: 170,997
United States
Message 1234491 - Posted: 21 May 2012, 12:48:17 UTC - in response to Message 1234475.

Thanks for the reply. I found the file, but have another question. I know my new 560's have 1.2 gig of ram, not sure about the 285, but I know the 260 doesn't have a ton of ram, so I would like to start out by having only the 2 560's doing multiple units. I read thru the file as best as I could but it appeared that the <count>1</count> settings were per app, not per card. Is it possible to spec that setting on each piece of hardware, or is it a global setting per application? Thanks!


If I understand the question, it's per application. If you tell it <count>.5</count> it will try to run two work units at a time on each piece of hardware.

Frankly, I don't know what happens if you try to run two WUs on a GTX 560 and two on a GT 240 (that can't handle it) in the same computer.

Profile HAL9000
Volunteer tester
Avatar
Send message
Joined: 11 Sep 99
Posts: 4424
Credit: 118,597,422
RAC: 132,747
United States
Message 1234497 - Posted: 21 May 2012, 13:07:54 UTC - in response to Message 1234491.

Thanks for the reply. I found the file, but have another question. I know my new 560's have 1.2 gig of ram, not sure about the 285, but I know the 260 doesn't have a ton of ram, so I would like to start out by having only the 2 560's doing multiple units. I read thru the file as best as I could but it appeared that the <count>1</count> settings were per app, not per card. Is it possible to spec that setting on each piece of hardware, or is it a global setting per application? Thanks!


If I understand the question, it's per application. If you tell it <count>.5</count> it will try to run two work units at a time on each piece of hardware.

Frankly, I don't know what happens if you try to run two WUs on a GTX 560 and two on a GT 240 (that can't handle it) in the same computer.

If the other card, in this case a GT 240, has enough memory it will run 2 at a time. Just much slower than makes it viable. Alternatively if the other card doesn't have enough memory there are a few things that can occur. None of them good.
____________
SETI@home classic workunits: 93,865 CPU time: 863,447 hours

Join the BP6/VP6 User Group today!

Profile ignorance is no excuse
Avatar
Send message
Joined: 4 Oct 00
Posts: 9529
Credit: 44,433,321
RAC: 0
Korea, North
Message 1234554 - Posted: 21 May 2012, 15:15:05 UTC - in response to Message 1234497.

Thanks for the reply. I found the file, but have another question. I know my new 560's have 1.2 gig of ram, not sure about the 285, but I know the 260 doesn't have a ton of ram, so I would like to start out by having only the 2 560's doing multiple units. I read thru the file as best as I could but it appeared that the <count>1</count> settings were per app, not per card. Is it possible to spec that setting on each piece of hardware, or is it a global setting per application? Thanks!


If I understand the question, it's per application. If you tell it <count>.5</count> it will try to run two work units at a time on each piece of hardware.

Frankly, I don't know what happens if you try to run two WUs on a GTX 560 and two on a GT 240 (that can't handle it) in the same computer.

If the other card, in this case a GT 240, has enough memory it will run 2 at a time. Just much slower than makes it viable. Alternatively if the other card doesn't have enough memory there are a few things that can occur. None of them good.

They'd be better off removing the 240 and just running 3 WU's at a time on the 560
____________
In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope

End terrorism by building a school

Profile HAL9000
Volunteer tester
Avatar
Send message
Joined: 11 Sep 99
Posts: 4424
Credit: 118,597,422
RAC: 132,747
United States
Message 1234557 - Posted: 21 May 2012, 15:22:37 UTC - in response to Message 1234554.

Thanks for the reply. I found the file, but have another question. I know my new 560's have 1.2 gig of ram, not sure about the 285, but I know the 260 doesn't have a ton of ram, so I would like to start out by having only the 2 560's doing multiple units. I read thru the file as best as I could but it appeared that the <count>1</count> settings were per app, not per card. Is it possible to spec that setting on each piece of hardware, or is it a global setting per application? Thanks!


If I understand the question, it's per application. If you tell it <count>.5</count> it will try to run two work units at a time on each piece of hardware.

Frankly, I don't know what happens if you try to run two WUs on a GTX 560 and two on a GT 240 (that can't handle it) in the same computer.

If the other card, in this case a GT 240, has enough memory it will run 2 at a time. Just much slower than makes it viable. Alternatively if the other card doesn't have enough memory there are a few things that can occur. None of them good.

They'd be better off removing the 240 and just running 3 WU's at a time on the 560

They could also run the 240 in a separate instance of BOINC, but that defeats the purpose of stuffing several GPU's in one box I think.
____________
SETI@home classic workunits: 93,865 CPU time: 863,447 hours

Join the BP6/VP6 User Group today!

Josef W. SegurProject donor
Volunteer developer
Volunteer tester
Send message
Joined: 30 Oct 99
Posts: 4299
Credit: 1,067,650
RAC: 1,014
United States
Message 1234571 - Posted: 21 May 2012, 15:48:31 UTC - in response to Message 1234475.

Thanks for the reply. I found the file, but have another question. I know my new 560's have 1.2 gig of ram, not sure about the 285, but I know the 260 doesn't have a ton of ram, so I would like to start out by having only the 2 560's doing multiple units. I read thru the file as best as I could but it appeared that the <count>1</count> settings were per app, not per card. Is it possible to spec that setting on each piece of hardware, or is it a global setting per application? Thanks!

As others have said, one setting applies to all cards using the application. In your upgrade thread, you posted these lines from BOINC startup:

5/20/2012 8:08:07 AM NVIDIA GPU 0: GeForce GTX 560 Ti (driver version 30124, CUDA version 4020, compute capability 2.0, 1280MB, 1665 GFLOPS peak)
5/20/2012 8:08:07 AM NVIDIA GPU 1: GeForce GTX 560 Ti (driver version 30124, CUDA version 4020, compute capability 2.0, 1280MB, 1665 GFLOPS peak)
5/20/2012 8:08:07 AM NVIDIA GPU 2: GeForce GTX 260 (driver version 30124, CUDA version 4020, compute capability 1.3, 896MB, 630 GFLOPS peak)
5/20/2012 8:08:07 AM NVIDIA GPU 3: GeForce GTX 285 (driver version 30124, CUDA version 4020, compute capability 1.3, 1024MB, 773 GFLOPS peak)


The GTX 260 has sufficient memory to run two tasks at a time, but pre-Fermi cards do not context switch very well so both the 260 and 285 might be less productive that way. The 560 Ti cards might more than make up the difference, though.

Given the ratio of the cards' speed ratings by nVidia's formula, you probably won't have any issues with tasks on the 260 running longer than allowed. But changing to two tasks per GPU might effectively increase the ratio, if it were over 10 there'd be that danger.
Joe

1 · 2 · 3 · Next

Message boards : Number crunching : Running multiple GPU tasks on one video card.

Copyright © 2014 University of California