CPU load issue with GPU3 Linux wrapper setup...

Moderators: slegrand, Site Moderators, PandeGroup

CPU load issue with GPU3 Linux wrapper setup...

Postby larryb63 » Sun Feb 03, 2013 6:28 am

Gang,

I noticed this problem with the GPU3 setup...and I'm wondering if anyone else has noticed the same.

I do folding on Slackware systems. Don't use any other Linux distro. I've been successful using the 64bit versions from 13.0 and beyond. These have the 32-bit Multilib setups as created by Alien Bob, one of the Slackware developers, along with his 32bit Wine builds....and it's working as expected.

What I've noticed is that when I setup a system to use the GPU3 wrapper version, that the FAH process ends up consuming an ENTIRE core on the PC. That means that a quad core only had three cores left to do other work with. A dual core has one left for other work.

I noticed this, since I run boinc (for WGC) on the CPU's, and run FAH on the GPU's.

While running the GPU3 wrapper with wine, the FOUR boinc tasks only use a total of 300% of a quad core. This leaves the FAH core with about 60-80% of IDLE time. I've been able to monitor this using 'top,' then hit '1' to show the specific percentages for each CPU core.

I thought it was a fluke.... some kernel misconfiguration, but through changing each part of the config...I have to pin it down to using the GPU2 wrapper with the cuda 2.3 toolkit that lets the entire CPU to be used. As soon as I switch up to the 3.0 cuda toolkit, I loose an entire core. I'm using different versions of wine (1.1.x, 1.4.x, and 1.5.x) and changing wine versions does not make any difference.

I even tried the 3.1 and 3.2 cuda toolkits just to see if that may make a difference. NO. Entire core lost to FAH. Idle percentages in the 60-80 range for the FAH core.

As soon as I backed the wrapper and toolkit to the 2.3 level...my CPU shows very little idle time. This is the only way I can get all of my CPU cores to operate near 100%.

I even tried different FAH client versions. 6.23, 6.30, and 6.41. No difference. They all run FahCore_11.exe.

May this have to do with the SLEEPWAIT variable within the GPU3 wrapper? I've tried to manipulate it, but it had no effect.

Could this be a problem (bug) in the cuda toolkit? I would believe that something is wrong there. I can't see how the wrapper dll could cause this kind of execution/scheduling oddity.

No exotic GPU's here either. 8800 GTS, 9800 GT, 9800 GTX+, and a GTX 260-216. No Fermi's or Kepler's.

Can folks out there take a look at your 'top' numbers, and see if the 3.0 wrapper is causing the same level of idle time on the FAH core?

--

Larry
larryb63
 
Posts: 30
Joined: Tue Oct 27, 2009 11:30 pm

Re: CPU load issue with GPU3 Linux wrapper setup...

Postby art_l_j_PlanetAMD64 » Sat Feb 09, 2013 1:54 pm

larryb63 wrote:Gang, I noticed this problem with the GPU3 setup...and I'm wondering if anyone else has noticed the same.

I do folding on Slackware systems. Don't use any other Linux distro. I've been successful using the 64bit versions from 13.0 and beyond. These have the 32-bit Multilib setups as created by Alien Bob, one of the Slackware developers, along with his 32bit Wine builds....and it's working as expected.

What I've noticed is that when I setup a system to use the GPU3 wrapper version, that the FAH process ends up consuming an ENTIRE core on the PC. That means that a quad core only had three cores left to do other work with. A dual core has one left for other work.

Can folks out there take a look at your 'top' numbers, and see if the 3.0 wrapper is causing the same level of idle time on the FAH core?

Larry

Hi Larry, I'm doing some GPU3/Linux/Wine testing on two Linux systems here, with fairly low-end CPUs:
  • AMD Athlon II X2 240 Dual-Core 2.8GHz, with a GTX 285 GPU
  • AMD Athlon 7550 Dual-Core 2.5GHz, with a GTX 460 'fermi' GPU
Everything is installed as described here, with these exceptions:
  • I'm using Debian v6.0.6 "squeeze" here, with
  • the NVidia v275.43 driver, and
  • the v6.41 GPU3 console client from the Folding@home-Win32-GPU_XP-641.zip file.

Doing this:
Code: Select all
ps aux | sort -k 3,3
on both systems gives this result:
  • AMD Athlon II X2 240 Dual-Core 2.8GHz, with a GTX 285 GPU
    • FahCore_a4 = 119%
    • FahCore_11 = 7.3%
  • AMD Athlon 7550 Dual-Core 2.5GHz, with a GTX 460 'fermi' GPU
    • FahCore_a4 = 179%
    • FahCore_15 = 16.5%
On the Athlon 240 system, I ran it this way:
Code: Select all
nice wine Folding@home-Win32-GPU.exe -verbosity 9 -forcegpu nvidia_g80

On the Athlon 7550 system, I ran it this way:
Code: Select all
wine Folding@home-Win32-GPU.exe -verbosity 9 -forcegpu nvidia_fermi

I found that the 'nice' was slowing down the Fermi GTX 460 to about a 9:00 TPF, which is terrible. Without the 'nice', the TPF drops to the 'normal' 3:45 for the P8073 WUs, on a stock-clock (675MHz) GTX 460.
art_l_j_PlanetAMD64
Over 1.04 Billion Total Points
Over 185,000 Work Units
Over 3,800,000 PPD
Overall rank (if points are combined) 20 of 1721690
In memory of my Mother May 12th 1923 - February 10th 2012
art_l_j_PlanetAMD64
 
Posts: 741
Joined: Sun May 30, 2010 2:28 pm

Re: CPU load issue with GPU3 Linux wrapper setup...

Postby larryb63 » Fri Feb 22, 2013 1:38 am

Art,

Thanks for checking! Much appreciated.

I determined when I first started doing GPU/Wine folding, that using 'nice' wasn't very nice in terms of ppd, so I run them all at normal priority. When I was testing this issue, I tried using nice again, and the TPF was going in the toilet, and PPD was still suffering. It would not change the CPU idle time, though.
larryb63
 
Posts: 30
Joined: Tue Oct 27, 2009 11:30 pm


Return to unOfficial Linux GPU (WINE wrapper) (3rd party support)

Who is online

Users browsing this forum: No registered users and 2 guests

cron