GPU computing not running

mickey
mickey
Joined: 18 Jan 10
Posts: 3
Credit: 24289
RAC: 0
Topic 195389

Hi guys, I am running into a very strange trouble.
Recently I have enabled Einstein to compute also on my nVidia Quadro FX 770M GPU. After this the client has started downloading one workunit of the application Arecibo Binary Pulsar Search (STSP) 1.11, but at the end of the download nothing has started. I have tried to suspend all other projects and workunits, but in this situation I have no application running, and the just downloaded workunit still "ready to start". My GPU computation is enabled, because before suspending other workunits I were running one app of primegrid using the GPU.
At last i saw that the client continue requesting new workunits for Arecibo 1.11: in this moment I have 18 WU Ready to start and nothing running.

From the prefs I have enabled "run always" and "use GPU always". This is my system config. Ask me for more details ;)

Gundolf Jahn
Gundolf Jahn
Joined: 1 Mar 05
Posts: 1079
Credit: 341280
RAC: 0

GPU computing not running

Your GPU doesn't have enough memory to process the Einstein CUDA tasks! (At least with the current drivers.)

Additionally, you seem to use an old BOINC version that doesn't handle that situation correctly. In the past, there have been several threads on this board about this topic.

If you can't free up enough GPU memory (by rebooting or using other drivers), your only option is to disable receiving GPU tasks from Einstein and to abort the already downloaded tasks.

There are some debug logging flags you could enable to show the amount of missing memory. In the aforementioned threads you can find more information on that topic.

Gruß,
Gundolf

Computer sind nicht alles im Leben. (Kleiner Scherz)

mickey
mickey
Joined: 18 Jan 10
Posts: 3
Credit: 24289
RAC: 0

RE: Your GPU doesn't have

Quote:
Your GPU doesn't have enough memory to process the Einstein CUDA tasks! (At least with the current drivers.)


wow, half a GB is not enough?

Quote:
Additionally, you seem to use an old BOINC version that doesn't handle that situation correctly.


I know, this is because I use the repository version of boinc, and for now its last version is 6.10.45

Quote:

If you can't free up enough GPU memory (by rebooting or using other drivers), your only option is to disable receiving GPU tasks from Einstein and to abort the already downloaded tasks.

There are some debug logging flags you could enable to show the amount of missing memory. In the aforementioned threads you can find more information on that topic.


Ok, I think that it is the only solution.
Thanks for the reply

mickey

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2143
Credit: 3001491932
RAC: 698360

RE: RE: Your GPU doesn't

Quote:
Quote:
Your GPU doesn't have enough memory to process the Einstein CUDA tasks! (At least with the current drivers.)

wow, half a GB is not enough?


Your card may be nominally 512MB, but your driver is only reporting 511MB.

If the Einstein project configuration has been set up to test, strictly, for exactly 512MB, then maybe you're failing the test, unecessarily. SETI had problems like that in the early days of CUDA.

If Einstein could relax the test by a couple of MB or so (which would make no difference - IIRC the app actually requires something in the low 400s), cards like yours could be brought into service.

mikey
mikey
Joined: 22 Jan 05
Posts: 12857
Credit: 1884348140
RAC: 281210

RE: RE: RE: Your GPU

Quote:
Quote:
Quote:
Your GPU doesn't have enough memory to process the Einstein CUDA tasks! (At least with the current drivers.)

wow, half a GB is not enough?

Your card may be nominally 512MB, but your driver is only reporting 511MB.

If the Einstein project configuration has been set up to test, strictly, for exactly 512MB, then maybe you're failing the test, unecessarily. SETI had problems like that in the early days of CUDA.

If Einstein could relax the test by a couple of MB or so (which would make no difference - IIRC the app actually requires something in the low 400s), cards like yours could be brought into service.

Hmmm I have this card "NVIDIA GeForce 9800 GT (499MB)" and it is crunching just fine. Now mine is a Win7 64bit machine and mickey is using Linux as his OS, there IS a difference, but the video card not reporting 512mb should not be it.

Gundolf Jahn
Gundolf Jahn
Joined: 1 Mar 05
Posts: 1079
Credit: 341280
RAC: 0

RE: Hmmm I have this card

Quote:
Hmmm I have this card "NVIDIA GeForce 9800 GT (499MB)" and it is crunching just fine. Now mine is a Win7 64bit machine and mickey is using Linux as his OS, there IS a difference, but the video card not reporting 512mb should not be it.


The problem isn't the reported total memory but the available memory.

And the older BOINC version would keep on downloading new GPU task because the total memory was sufficient, but it would never start them because the lack of available memory.

Gruß,
Gundolf

Computer sind nicht alles im Leben. (Kleiner Scherz)

Mike Hewson
Mike Hewson
Moderator
Joined: 1 Dec 05
Posts: 6592
Credit: 331300566
RAC: 304685

RE: Hmmm I have this card

Quote:
Hmmm I have this card "NVIDIA GeForce 9800 GT (499MB)" and it is crunching just fine. Now mine is a Win7 64bit machine and mickey is using Linux as his OS, there IS a difference, but the video card not reporting 512mb should not be it.


Some machine setups I've seen ( laptops ) can also have a mapping/overlap b/w system and graphics memory. So there may be either double counting, or robbing one to pay the other. Thus the phrase 'video card memory' and 'video memory' aren't equivalent alas. In any case, at the time of allocation of graphics resources to crunching it needs to be sufficiently free of it's primary obligation as an interface device.

What is, say, the absolute minimum with a screen's worth of 1024 x 768 @ 24 bits per pixel? That comes to 18874368 bits. Big deal. But of course it aint a one-to-one in real life. So what is it the cards normally do with their memory beyond the above bland bitmap calculation I used to do in the MS-DOS/8086 days? I've seen mentioned vertex arrays, textures, display lists etc in my OpenGL reading. A quicker cache idea in a server side model? A huge parking lot for inputs and outputs with polygon algorithms?

Cheers, Mike.

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

... and my other CPU is a Ryzen 5950X :-) Blaise Pascal

Jord
Joined: 26 Jan 05
Posts: 2952
Credit: 5893653
RAC: 0

RE: Some machine setups

Quote:
Some machine setups I've seen ( laptops ) can also have a mapping/overlap b/w system and graphics memory.


BOINC only counts the memory actually on the chip, not any additional memory you can add to it through system memory. Best seen on Dell systems, which comes with a 128MB videocard, which can be added to from system memory, up to 1GB. They will show as 128MB cards to BOINC only, no matter how much the owner added through BIOS.

mickey
mickey
Joined: 18 Jan 10
Posts: 3
Credit: 24289
RAC: 0

By now I remember that in the

By now I remember that in the past this pc has crunched some WU on my GPU. Sincerly I don't remember if it was version 1.11 or earlier, but I am sure that I has completed successifully some WU.
So it is strange that by now it won't run any WU on GPU. Maybe i was running them under windows and not Linux (I can't remember).

Anyway the video driver is the last providen by nVidia for my video card and Linux, and my pc is a Dell Precision M4400, from the site I have read that the video memory is full dedicated to the GPU and not shared on the RAM.

Comment viewing options

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