Block 1 GPU, Run on the Other

Message boards : Number crunching : Block 1 GPU, Run on the Other

Author	Message
Karplusan Send message Joined: 11 Mar 09 Posts: 10 Credit: 3,152,021 RAC: 0 Level Scientific publications	Message 17848 - Posted: 3 Jul 2010 \| 8:34:17 UTC
	Is there a way to tell GPUGRID to only run tasks on a specific video card? Or tell it to not run on specific cards? I currently have two different video cards capable of running BOINC cuda apps. A GeForce GTX 470 and a GeForce 8600 GTS. The GeForce GTX 470 runs GPU Grid great and can finish 2-3 WU's in a day if I let do so, and no other projects took it over. The 8600 GTS could finish one in 32 hours, if it did not run into an Error while Computing, which is what it's done with every GPUGRID task it's been given since 2008. The 8600 is still a good CUDA card and runs other BOINC cuda apps great, but it is wasting time with GPUGRID. Is there a way to make sure that card ignores GPUGRID and only runs other apps?
	ID: 17848 \| Rating: 0 \| rate: / Reply Quote

ExtraTerrestrial Apes Volunteer moderator Volunteer tester Send message Joined: 17 Aug 08 Posts: 2705 Credit: 1,311,122,549 RAC: 0 Level Scientific publications	Message 17875 - Posted: 4 Jul 2010 \| 20:34:28 UTC - in response to Message 17848.
	BOINC is supposed to do everything automatically for you, so it doesn't provide many means to let you tell it, what you actually want it to do (apart from ressource share). There's this "use all GPUs" flag. If it's disabled (default) BOINC refuses to use GPUs with lower specifications (not sure about the current rules). That probably wouldn't let you run other projects on the smaller card, though. Then there's a server setting to tell BOINC which CUDA hardware compatibility level a project requires. The minimum for GPU-Grid is 1.1, which your 8600 meets and the 470 exceeds. One could probably hack BOINC here and trick it into thinking that GPU-Grid wouldn't run on the 8600, but other projects still would. Ideally that's just adding one single "if" cluase at the right place, but would have to be repeated for every new BOINC version (as this likely wouldn't find its way back into the main code base). It's been a while since the last time I cried for ressource based project allocation as an advanced option to supplement the ressource share based scheduling. That could handle such cases easily. MrS ____________ Scanning for our furry friends since Jan 2002
	ID: 17875 \| Rating: 0 \| rate: / Reply Quote

skgiven Volunteer moderator Volunteer tester Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level Scientific publications	Message 17880 - Posted: 5 Jul 2010 \| 9:37:58 UTC - in response to Message 17875.
	Take the 8600 GTS out of the system and put it into another system! You could try using it at Folding@home, by (as MrS said) adding the following line to the cc_config.xml file, <use_all_gpus>0</use_all_gpus> I'm guessing this will allow GPUGrid to use the Fermi, but it might just see the 8600, if that is in the first PCIE slot; so you might have to swap them over. Folding is a non-Boinc program, so you may have to ask a few questions over at folding. It's a very big project, so you should get some support. They only have a command interface for Fermi, so their standard GPU-app should just see the 8600. I think the 8600 GTS still works there, but no guarantees. Let us know how you get on,
	ID: 17880 \| Rating: 0 \| rate: / Reply Quote

Karplusan Send message Joined: 11 Mar 09 Posts: 10 Credit: 3,152,021 RAC: 0 Level Scientific publications	Message 18375 - Posted: 20 Aug 2010 \| 3:23:23 UTC - in response to Message 17848.
	By Default, BOINC had disabled the "smaller" GPU. I had to add the line to the cc_config.xml file to even get BOINC to recognize it. After running with that configuration a few days, I posted my original message. Other projects were giving me a similar error. SETI would download cuda_Fermi units, which the Fermi card would plow through, and the slower one would fail. My solution has been inelegant. It's now been near 2 months, and I don't see so many failed tasks anymore. I'm guessing there have been enough failed Work Units reported that somebody has come up with a fix, on a per project basis. No, putting the card in another machine is not an option. There is only 1 motherboard in my house that has PCI-E slots, let alone PCI-E 16. On top of that, the 8600 has PhysX, which does help games look and run better, so it's not a total loss sitting in the same case. I bought these cards to play games. I am also away from computer a lot and figure the potential number crunching is high. With 2 GPU's, I figure that potential is higher. Thanks for the feedback. I'm away from computer too much, and my own response time can be rather slow.
	ID: 18375 \| Rating: 0 \| rate: / Reply Quote

Message boards : Number crunching : Block 1 GPU, Run on the Other

//