CUDA performance / GPU Utilisation

Questions and Answers : Unix/Linux : CUDA performance / GPU Utilisation
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile ML1
Volunteer moderator
Volunteer tester

Send message
Joined: 25 Nov 01
Posts: 20084
Credit: 7,508,002
RAC: 20
United Kingdom
Message 946909 - Posted: 13 Nov 2009, 12:49:35 UTC
Last modified: 13 Nov 2009, 12:50:33 UTC

So, I've CUDA working for s@h.

So, how do you see how effective it is?

Are there any utilities to show GPU utilisation?

I'm seeing some effect in that the GPU temperature varies between 40 deg C and 50 deg C. Otherwise?...

Also, I have vsync sync checked so that display applications will max out at 60 fps rather than silly 1000's fps. Is that correctly ignored for CUDA?

(Experimenting, I see no difference but then how do you immediately tell?)


For the s@h tasks, the CPU utilisation is surprisingly low at less than 1% of one CPU core. Or is that actually a "surprisingly good"?!

Happy fast crunchin',
Martin
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
ID: 946909 · Report as offensive
Profile skildude
Avatar

Send message
Joined: 4 Oct 00
Posts: 9541
Credit: 50,759,529
RAC: 60
Yemen
Message 946931 - Posted: 13 Nov 2009, 15:26:59 UTC - in response to Message 946909.  

Effective? It completes WU's very quickly compared to the CPU. I guess that makes it effective. check your computer on your account page. the rac should be increasing. You might be able to adjust a system monitor to check your GPU usage. not sure since I dont know which flavor you are using

I would assume that Nvidia might have put a console out for linux so you could manipulate your video card settings.

The GPU should increase a great deal. I recall most folks running around 60C

The GPU should only require the CPU to load the WU to the GPU then sit idle until the WU finishes. I'm assuming what you are seeing is normal. It works the same way on windows



In a rich man's house there is no place to spit but his face.
Diogenes Of Sinope
ID: 946931 · Report as offensive
Profile ML1
Volunteer moderator
Volunteer tester

Send message
Joined: 25 Nov 01
Posts: 20084
Credit: 7,508,002
RAC: 20
United Kingdom
Message 947943 - Posted: 18 Nov 2009, 0:42:58 UTC - in response to Message 946931.  
Last modified: 18 Nov 2009, 0:44:07 UTC

Effective? It completes WU's very quickly compared to the CPU. I guess that makes it effective. check your computer on your account page. the rac should be increasing. You might be able to adjust a system monitor to check your GPU usage. not sure since I dont know which flavor you are using...


Well, the RAC does indeed appear to be rapidly increasing.

Curiously however, the GPU lingers for long periods near it's baseline temperature of 41 deg C, often meanders around 44 deg C, and only briefly and occasionally hits near the max I've ever noticed at about 50 deg C.

Are any of the lunatics toying with pipelining multiple WUs through but phased so that the GPU is more highly and more consistently utilised?... ;-)

The nVidia console shows the fan speed and temperature, and I have the "power mizer" set to "performance".

The system tools has some very nice displays to show all manner of system stuff including some very pretty graphs for just about everything except the GPU.

So... There appears to be no obvious tool to show what utilisation is being achieved for the GPU. Other than to try to deduce something from the WU RAC...


Aside:

In the s@h app_info.xml, there is:

<plan_class>cuda</plan_class>
<avg_ncpus>0.040000</avg_ncpus>
<max_ncpus>0.040000</max_ncpus>


How is the ncpus used? They were set to "0.44". I've just made a guess towards what I actually see from "top" whereby s@h is taking less than 1% CPU time.


Happy crunchin',
Martin

Mandriva 2010 x86_64
Boinc 6.10.17
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
ID: 947943 · Report as offensive
Profile ML1
Volunteer moderator
Volunteer tester

Send message
Joined: 25 Nov 01
Posts: 20084
Credit: 7,508,002
RAC: 20
United Kingdom
Message 948085 - Posted: 18 Nov 2009, 19:00:46 UTC

The s@h GPU app is now laying idle for periods after having run continuously for many hours.

I've got s@h on a small resource share... Can the GPU get starved and go idle due to the resource shares?!


Happy crunchin',
Martin

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
ID: 948085 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 948089 - Posted: 18 Nov 2009, 19:22:04 UTC - in response to Message 948085.  

Yes, same way as your CPU can go idle on a small resource share. There's only one resource share, one DCF, one debt per project. So when there's now debt gathered by the GPU, BOINC will do other projects first until that debt is gone.
ID: 948089 · Report as offensive
Profile ML1
Volunteer moderator
Volunteer tester

Send message
Joined: 25 Nov 01
Posts: 20084
Credit: 7,508,002
RAC: 20
United Kingdom
Message 948095 - Posted: 18 Nov 2009, 20:28:05 UTC - in response to Message 948089.  

Yes, same way as your CPU can go idle on a small resource share. There's only one resource share, one DCF, one debt per project. So when there's now debt gathered by the GPU, BOINC will do other projects first until that debt is gone.

Except...

In this case, there are no other projects using the GPU. Also, s@h's share of the CPU is set at about x5 or much more than the actual CPU time being used...

So the GPU goes idle if other projects do not use up their GPU share?

Happy crunchin',
Martin

See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
ID: 948095 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 948110 - Posted: 18 Nov 2009, 21:35:36 UTC - in response to Message 948095.  

Except..

In this case, there are no other projects using the GPU.

That doesn't matter. You only have one debt, which is used for all, be it if you use only GPU, only CPU, or both. Go into debt as opposed to the other projects you are attached to and Seti will not request more work until that debt is gone.

So you probably still have work for the CPU, it just won't request any more work at this moment for either.
ID: 948110 · Report as offensive

Questions and Answers : Unix/Linux : CUDA performance / GPU Utilisation


 
©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.