Advanced search

Message boards : Number crunching : Just a few Linux/GPU questions

Author Message
brinktastee
Send message
Joined: 19 Feb 10
Posts: 2
Credit: 2,602,823
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwat
Message 19964 - Posted: 17 Dec 2010 | 0:01:00 UTC

I'm new to the GPU crunching thang and I have a few questions.
1. What does this message mean? GPUGRID Message from server: Fermi-class GPU not supported by cuda2.2
2. Can I crunch more than one work unit at a time with one card.
3. How can I find the GPU utilization in Linux. I notice others posting there utilization here.
4. In Linux is there a way to control the fan speed on the GPU?
5. How do I find out what version of cuda I'm running in Linux?

Thank you for your time.
Brink
-->Ubuntu 10.10/boinc 6.10.58/EVGA GTX 470 SC/Driver 260.19.29<--
____________

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,961,729,435
RAC: 299,440
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 19970 - Posted: 17 Dec 2010 | 1:12:26 UTC - in response to Message 19964.
Last modified: 8 Jun 2011 | 11:03:18 UTC

I'm new to the GPU crunching thang and I have a few questions.
1. What does this message mean? GPUGRID Message from server: Fermi-class GPU not supported by cuda2.2
2. Can I crunch more than one work unit at a time with one card.
3. How can I find the GPU utilization in Linux. I notice others posting there utilization here.
4. In Linux is there a way to control the fan speed on the GPU?
5. How do I find out what version of cuda I'm running in Linux?

Thank you for your time.
Brink
-->Ubuntu 10.10/boinc 6.10.58/EVGA GTX 470 SC/Driver 260.19.29<--


Hi Brink, I'm no Linux expert but I will try to help.

1. I guess you might have needed to restart your X-Server after installing the driver. Perhaps you have done this now? That driver should come with the latest CUDA support; up to 3.2. The message "Fermi-class GPU not supported by cuda2.2" is related to CPU requests; the project does not run cpu only tasks.

2. Yes, but you might not want to concentrate on getting it to work first, and if you work out how to use swan_sync=0 on Linux and free up one CPU core/thread in Boinc (use 75% CPU for example on a quad) you might be happy with 90% GPU utilization.

3. On Win we just use GPUZ or EVGA Precision. Not sure but I think I found some command tool for Ubuntu 10.10 but I had to keep refreshing it, it was not live.

4. Yes, Liveonc found a way and posted about it some time ago.
Quote, "To manually adjust the fan speed (to prevent overheating & increase the chances that WU's don't fail due to high temps). Install via Synaptic, Nvclock & write in terminal $sudo nvclock -f -c 1 -F 99 & ($sudo nvclock -f -c 2 -F 99 if you have a second GPU, 3 if you have three GPU's & so on). "

5. Try Boinc Manager, Messages, 13th line down?


Hope that's of some use,

-
Not recommended but Nibitor and nvflash (with boot files) can be used to modify the clocks/flash some cards - there's some Youtube vids on this.

brinktastee
Send message
Joined: 19 Feb 10
Posts: 2
Credit: 2,602,823
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwat
Message 19985 - Posted: 18 Dec 2010 | 0:22:07 UTC - in response to Message 19970.

Thank you skgiven for the quick response.
1. I still get that message even after a reboot.Fri 17 Dec 2010 05:59:21 PM CST GPUGRID Message from server: Fermi-class GPU not supported by cuda2.2 But, I am getting work units so that's good. Should I just ignore it?
2. I'm guessing you want me to put this in my .bashrc in my home directory--> export SWAN_SYNC=0 And only run 7 of my 8 threads in boinc. Done and done.
3. I googled around for GPU utilization on Ubuntu and so far nothing. If you find your command that would be cool if you post it.
4. I played with this for an hour and could not see any change. Oh well.
5. Found it on line 15 :) I have 2 more lines for my GUI RPC stuff. Curious why the driver is unknown though....?
NVIDIA GPU 0: GeForce GTX 470 (driver version unknown, CUDA version 3020, compute capability 2.0, 1279MB, 1120 GFLOPS peak)

Anywho thanks for your help, much appreciated!


____________

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,961,729,435
RAC: 299,440
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 19986 - Posted: 18 Dec 2010 | 2:06:44 UTC - in response to Message 19985.
Last modified: 18 Dec 2010 | 2:11:34 UTC

Hi brink,


1. I still get that message even after a reboot.Fri 17 Dec 2010 05:59:21 PM CST GPUGRID Message from server: Fermi-class GPU not supported by cuda2.2 But, I am getting work units so that's good. Should I just ignore it?

Yes, you are getting tasks and they are completing. As for the message, it’s probably just the server working out that you have a Fermi and to send the 6.13app rather than the 6.12app; the analogue of the "GPUGRID Message from server: CUDA version 3.1 needed" message.

2. I'm guessing you want me to put this in my .bashrc in my home directory--> export SWAN_SYNC=0 And only run 7 of my 8 threads in boinc. Done and done.

Yes, well done for working out how. To see if the setting stuck in the right place type, echo $swan_sync
It should return "0"

3. I googled around for GPU utilization on Ubuntu and so far nothing. If you find your command that would be cool if you post it.

I thought I had used a command line reporting tool to report the GPU utilization once, but it was not live. The system has since been rebuilt with Win and the only rough notes I have relate to cat /proc/cpuinfo, dmidecode, nvclock (might not work for Fermi's) and perlmon. Perlmon has a Video card info tab, so perhaps I was accessing that from a prompt, but I'm not sure.

4. I played with this for an hour and could not see any change. Oh well.

Yeah, I found 10.10 difficult; had to use unfamiliar commands like “sudo nautilus ...” for MM security reasons. I had used earlier Ubuntu and Kubuntu versions for a bit, so it was back to square one with MM. Kept having to install things to install other things. You even have to setup Ctrl+Alt+F7 just to restart the X-server quickly. I had messy driver/setup issues and could not get a simple workaround. Ended up at a black screen after an update attempt, so left it at that. I had one GTX260 and one GT240. But things should be much easier with a single Fermi; no cc_config.xml file to build.
Who knows, maybe this no longer works with MM, but it was install coolbits, and from the command type, sudo nvclock –f –c 1 –F 75
The –F 75 is supposed to set the fan to 75%

5. Found it on line 15 :) I have 2 more lines for my GUI RPC stuff. Curious why the driver is unknown though....?
NVIDIA GPU 0: GeForce GTX 470 (driver version unknown, CUDA version 3020, compute capability 2.0, 1279MB, 1120 GFLOPS peak)

Just another driver reporting issue, not really important. Seen that a lot over the last year. Should say 260.19.29 but the rest of the line is correct.

The KASHIF_HIVPR task you have should be faster.


Good luck,

Rabinovitch
Avatar
Send message
Joined: 25 Aug 08
Posts: 142
Credit: 34,064,323
RAC: 46,445
Level
Val
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 20073 - Posted: 30 Dec 2010 | 8:33:32 UTC - in response to Message 19985.


3. I googled around for GPU utilization on Ubuntu and so far nothing. If you find your command that would be cool if you post it.

Try to visit this forum. They say there that nvidia-smi -a command should return nice and informative data about GPU and video memory utilization.

____________
From Siberia with love!

Kenneth Larsen
Send message
Joined: 11 Feb 09
Posts: 6
Credit: 162,131,296
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 20962 - Posted: 14 Apr 2011 | 19:01:17 UTC
Last modified: 14 Apr 2011 | 19:02:27 UTC

As noted above, 'nvidia-smi -a' does indeed show the gpu and memory usage. It should be part of the nvidia-drivers package.
I'm currently trying to integrate it in the conky desklet to easily monitor it.

EDIT: Sorry for reviving such an old post, I hadn't seen the date...

Post to thread

Message boards : Number crunching : Just a few Linux/GPU questions