Is there any solution available to make GPUGRID function with fermi cards in Linux environement?

All Wus i got in linux (6.04) crashed. I tried with Ubuntu/Xubuntu/Kubuntu and Suse in their last x64 flavours.

I'm a linux newbie as i just want to use it for crunching faster.

I'm not sure they have released the linux version for fermi yet, should be any day now. To make sure you get the correct WUs for your card and OS go into your GPUGrid preferences and make sure you have "Yes to Run test applications" and then select "ACEMD" and "BETA" in the Run only the selected applications section.

---

Thanks - Steve

I'm a N00B to, it's limited what I can do to help, but being a N00B makes it easier to understand other N00B's (I guess).

Here's a guide to get the latest Drivers for Linux, which aren't always available the easy way.

1) First step: you must reset the Xorg to its default conf. Before, you should backup it, to avoid any mistakes.

In the Terminal:

$ sudo cp /etc/X11/xorg.conf /etc/X11/xorg.conf.original
$ sudo dpkg-reconfigure -phigh xserver-xorg

2) Installing packages and dependencies:

$ sudo apt-get install build-essential linux-headers-`uname -r`

3) Remove old version drivers:

$ sudo apt-get --purge remove $(dpkg -l | grep nvidia | awk '{print $2}')

Download the driver from nvidia website. You must check your system architecture (x86, 64...)

4) In my case: Ubuntu 64:

$ wget ftp://download.nvidia.com/XFree86/Linux-x86_64/195.36.24/NVIDIA-Linux-x86_64-195.36.24-pkg2.run -O NVIDIA-Linux-195.pkg.run

5) Now, move the installer to /usr/src and link it. Follows:

$ sudo install NVIDIA-Linux-195.pkg.run /usr/src/
$ sudo ln -s /usr/src/NVIDIA-Linux-195.pkg.run /usr/src/nvidia-driver

Kill X:
6) Time to stop X on GDM. So, press "Ctrl+Alt+F1", login and stop gdm:

$ sudo /etc/init.d/gdm stop (If KDE is used write instead: $ sudo /etc/init.d/kdm stop)

Installing NVIDIA Driver
7) Installing:

$ sudo sh /usr/src/nvidia-driver

8) When it is done, restart your computer:

$ sudo reboot

After boot up, go to terminal and:

$ sudo nvidia-xconfig
(I didn't need to do this, but this step is in the origial step-by-step)

BTW, I got this from here: http://forums.nvidia.com/index.php?showtopic=99513 & changed a little bit.

To search for the latest Nvidia Driver, go to http://ftp://download.nvidia.com/XFree86/ to find what they have.

---

I think Snow Crash has it right, there has been no Linux specific Fermi app released yet. I know I myself am checking the site many times a day waiting for it. I'm dying to see what my 470s can do vs my 295. I never crunched wu's on windows with my 295 so I have no basis for comparison until they release the Linux app.

I think the gpugrid guys use Linux in the lab, hopefully we'll see the Linux app soon, but so far they have mentioned almost nothing about it whatsoever. I think a lot of development time has been devoted to ATI lately; I really hope that doesn't mean us Linux guys will just have to wait...

Thank you all for useful informations. I'm going to wait a bit before crunching under Linux, i'm glad to discover new OS via GPUGRID participation.

Again thx guys.

PS: Do you think Open suse 11.2x64 is a good distro to crunch and learn about linux world? This one is the one i prefer after testing Ubuntu and derivated included Mint, Mandriva and Fedora. What is your advice?

6.04 tasks are not for Fermi.

Thank you all for useful informations. I'm going to wait a bit before crunching under Linux, i'm glad to discover new OS via GPUGRID participation.

Again thx guys.

PS: Do you think Open suse 11.2x64 is a good distro to crunch and learn about linux world? This one is the one i prefer after testing Ubuntu and derivated included Mint, Mandriva and Fedora. What is your advice?

SuSE is a great distro, very robust and lots of development. I personally use Ubuntu, mostly because the deb package management system is fantastic.

Finally, I opted for the Ubuntu distro, mainly due to install via wubi which allows Windows to coexist just seven and Linux without having to suffer the burden of Grub 2.

My first workunits were calculated in 425 seconds or less, I'm pretty happy with the results.

Wubi is really an extraordinary idea for a setup like mine, again, everything here has installed a ssd, I did not suffer loss of velocity on my disk access, the best of both worlds ...


(Google traduction from French, sorry if it's not accurate)

My machine has been happily crunching WUs all afternoon, no errors, everything completed and validated.

431secs is the slowest one i've seen. This is almost twice as fast as the few beta units I crunched while booted into win7.

I have the problem after reboot where my gpu is not found, then i found a workaround with two shortcuts on my desk, one to stop client, the other to restart it.

No "panache" here but it does the job.

I think increasing the start up delay instead of manually stopping and starting would work.

<cc_config>
<options>
<start_delay>60</start_delay>
</options>
</cc_config>

---

Thanks - Steve

It will indeed be interesting to see if that works, but I think it failed last time it was tried.

IIRC, the 'start_delay' in cc_config only controls how long the science apps wait, after BOINC itself has already started (BOINC has to be running, because it's BOINC that reads cc_config).

The trouble desribed here happens if BOINC starts before the CUDA driver is ready - BOINC has to detect the CUDA card through its drivers before it can schedule CUDA science apps properly.

I think there have been some delayed-load BOINC scripts posted around the place, but I'm afraid I haven't got a link handy.

Thanx Snow Crash! Works for me. Just don't put it on the startup programs list, because then you have to stop & start, but otherwise the delay does the trick ;-)

---

Does not work for me but a launcher restarting boinc client does actually the job.

Xubuntu permit to compute really faster, according to my personal tests, than Ubuntu or kubuntu. Give it a try.

You're right, I was a bit too optimistic. It worked for me, then it stopped working. I restarted my pc several times & it seemed as if the problem was gone, but then it came back & never worked again. I guess restarting BOINC is the only way to go on Linux.

I've got lots of performance on Ubuntu & Kubuntu 10.4, only problem is, I've never had so many failed WU as I now have. I think it's heat though, Linux WU's seem to be far more aggressive then Windows WU's. I don't think it's Ubuntu, because even my Mint linux 8 pc's Ubuntu 9.10 started to have them. I don't think it's Ubuntu, Nvidia drivers, or Boinc, but rather the GPUGRID WU's. Summer hasn't been that hot here & sometimes with the windows open 24/7, it's even colder then it was during winter, i think it's something recent.

---

For Information of linux users, to create proper launcher that stop and restart Boinc when "gpu missing" appears in its messages lines, the command line is :


sudo /etc/init.d/boinc-client restart


Don't forget to execute it in a terminal or it could not work as wanted.


Here is the stats i get crunching under Xubuntu :

<core_client_version>6.10.17</core_client_version>
<![CDATA[
<stderr_txt>
# Using device 0
# There is 1 device supporting CUDA
# Device 0: "GeForce GTX 470"
# Clock rate: 1.22 GHz
# Total amount of global memory: 1341325312 bytes
# Number of multiprocessors: 14
# Number of cores: 112
MDIO ERROR: cannot open file "restart.coor"
# Time per step (avg over 50000 steps): 8.179 ms
# Approximate elapsed time for entire WU: 408.944 s
08:07:37 (3004): called boinc_finish

</stderr_txt>
]]>


Average 8ms/step with stock gtx 470 and Core 2 quad Q6600 runing 3.6Ghz 450 Fsb X 8 watercooled.

90°C for gpu temp 38°C for hotest core of Cpu at ambiant 27°C.

No crash since new Wus a couple of days. All runs fine for me here, yesterday i reached the 120 wus/day quota near 12H crunching since 00H00. Fermi works well on linux x64!!!

Hello all.

In my case I always calculate an error with my gtx480.
I installed the latest nvidia driver, the latest stable version of boinc and nothing I have done those silly mistakes

Perhaps this is because you put a GTX285 a Tesla C1060 and a GTX480 all into the same system.

The GTX480 will only crunch CUDA30 tasks, and the other cards will only crunch cuda tasks.

When your GTX480 picks up a cuda task it will fail, then get another task and fail, and so on until it randomly gets a CUDA30 task. The same applies for the Tesla and GTX285, except they will fail cuda30 tasks and crunch cuda tasks.

Put the Fermi into a different system.

That was it, thank you

After the disappointing experience with the poor new Win 7 drivers for Fermi (257.xx, which is even worse than the 197.xx driver under Win 7) I tried to run GPUGrid on Ubuntu 10.04.
Results for my GTX470:
7 out 8 work units crashed
1 work unit took 36,791.16 seconds for a 6803 credits task: (http://www.gpugrid.net/workunit.php?wuid=1665624). This took place on a i7-920 on stock speed equipped with one GTX470 stock clocked, BOINC 6.10.17 with setting for 7 cores for CPU crunching in order to have one core for the GPU.
I recognized that the task ran with a CPU usage of about 5% - rather low for a Fermi.

The unit that is running on the machine right now is also extremely slow: After 37 min the progress bar shows 5% - even a GTS 250 under Win 7 is faster. With this speed the GTX470 generates about 16-17k Credits/day :-D

I installed the latest driver 256.35 exactly as decribed in this posting: http://www.gpugrid.net/forum_thread.php?id=2150&nowrap=true#17010

Any ideas what went wrong?