SETI MB CUDA for Linux

[Jason G]

Holy €^#!  Nez has a serious problem!

Eric's FAQ on the subject, in the staff blog, at http://setiathome.berkeley.edu/forum_thread.php?id=56450

[dmol]

Hi All,
I'm trying to get CUDA working on a Quadro NVS 295 without much success. Its failing back to CPU processor due to lack of memory. I know its not a great card but I want to utilise all 4 cpu cores as well as the GPU. Is it even possible to get this card working due its lack of memory?:

OS:                              Ubuntu 9.10
Kernel:                        2.6.31-16-generic x86_64
CPU:                           Intel(R) Xeon(R) CPU   (quad)  E5506  @ 2.13GHz
Memory:                     4Gb
Nvidia card:               Quadro NVS 295 256Mb RAM
Nvidia-drivers:           Linux x86_64 190.53 (latest) including CUDA v2.3 driver
Nvidia-toolkit:            v2.3
Xwindow running:    Yes
Boinc Version:          6.10.17 for x86_64-pc-linux-gnu
Optimised app:        SETI@home MB CUDA_2.2 608 Linux 64bit SM 1.0 - r12 by Crunch3r VLAR autokill mod

The card is detected fine by boinc:
18-Dec-2009 16:51:12 [---] NVIDIA GPU 0: Quadro NVS 295 (driver version unknown, CUDA version 2030, compute capability 1.1, 255MB, 21 GFLOPS peak)
18-Dec-2009 16:51:12 [SETI@home] Found app_info.xml; using anonymous platform

Sample task:  http://setiathome.berkeley.edu/result.php?resultid=1452918416 (Note: Did not use the "VLAR autokill" version here but result the same)

<stderr_txt>
setiathome_CUDA: Found 1 CUDA device(s):
   Device 1 : Quadro NVS 295
           totalGlobalMem = 267714560
...
setiathome_CUDA: CUDA Device 1 specified, checking...
   Device 1: Quadro NVS 295 is okay
SETI@home using CUDA accelerated device Quadro NVS 295
setiathome_enhanced 6.01 Revision: 402 g++ (GCC) 4.1.2 20070925 (Red Hat 4.1.2-33)
libboinc: BOINC 6.5.0
....
Cuda error 'cudaMalloc((void**) &dev_GaussFitResults' in file './cudaAcceleration.cu' in line 317 : out of memory.
setiathome_CUDA: CUDA runtime ERROR in device memory allocation (Step 1 of 3). Falling back to HOST CPU processing
</stderr_txt>



SETI preferences:
Use cpu:         yes
Use GPU:       yes
Applications:  SETI@home enhanced only

App_info.xml main information:
  <name>setiathome-6.08.CUDA_2.2_x86_64-pc-linux-gnu</name>
<avg_ncpus>3.5</avg_ncpus>
<max_ncpus>3.5</max_ncpus>

There is no cc_config file in use.

 

[riofl]

since you share this card with your desktop i seriously doubt it will work. cuda needs a minimum of 256mb vidram for itself which you do not have if your desktop uses some of it. if yiou are not critical of your graphics capabilities for your desktop i recommend an inexpensive card for just your desktop and let this 295 handle cuda. some have managed to smooge enough to make a single card work but to be blunt i doubt it is worth all the hoops to jump through.

[dmol]

Thanks riofl,
I take it that there's no way to configure the SETI app to use less memory?
Can anyone recommend any other project that will work on this card/OS without using the CPU?

[Claggy]

Collatz has Linux Cuda apps, it's especially good for GPU's with little memory,
My laptop's 128Mb 8400m GS works fine (on Windows_X86), no lag, and 0 to 1% to feed the GPU.

Claggy

[sunu]

dmol, you can try the client from http://calbe.dw70.de/linux64.html . It is release r06 and it is a bit less memory hungry.

[riofl]

i am going to try cuda 2.3 again to iron out any oddities before my project begins next year...  i cannot find a cuda 2.3 application though.. the one i have now is

setiathome-6.08.CUDA_2.2_x86_64-pc-linux-gnu

any clues where to look? i can imagine one must have been generated by now.

also have the newer kernels fixed the slower crunching time reported previously? i am still running 2.6.29. stay with it?

[sunu]

There are no 2.3 cuda applications. Recompiling the apps with 2.3 didn't show any speed-up. All speed-up gains are from the 2.3 cuda libraries and the nvidia drivers not the apps themselves.

As for kernel versions you can always try new ones. I've never done a speed comparison among different kernel versions. I've switched to sidux and right now I'm using 2.6.32-2, the latest stable official, and don't see any big difference but as I said I've never done a speed comparison.

[riofl]

ok cool. thanks.

question. do you know what VLAR angle the vlar killer app uses to trigger rejection? i have been using that cpugpu script and recently i have had a slew of computation errors and when i look at them they all were slightly higher than 0.13 such as 0.133 so they got fed to the gpu and got rejected as vlar.  i am trying 0.14 in the script now.

[sunu]

Unfortunately I don't know the range of the VLAR kill.

[riofl]

ok no prob. trial and error works well   seems 0.14 was a huge improvement but still gave a few so i moved it to 0.15 and that seems to do it so far. it will probably give the cpu a couple that the gpu could be happy with but.....

[sunu]

If you have a high output machine, even with 3 gpus like I do, the script isn't much useful. VLARS can be so many that the CPU won't be able to do them in time so you'll have to abort (at least some of) them anyway. I tried to use it for a week and my CPU's cache kept growing and growing out of proportions.

So I just do a search for AR 0.19 or lower and abort them. I don't leave them to error out with VLAR-kill so I can differentiate them in my error tasks page, VLARs (aborted) and true errors that need attention.

[riofl]

interesting... i probably wont care much once i replace this tesla. it locks solid on vlars but works ok on the rest. i was waiting for the gtx 295 to go down but it seems to have increased in price instead.

[sunu]

All cards suffer from VLARs, it is the software not the hardware at fault.

As for GTX295s, their price may have gone up because of their reduced production numbers. Still they are cheaper than when I bought mine last July-August.

[riofl]

yeah. ill probably just bite the bullet and get one.

vlars kill the tesla. it locks up hard and i have to stop and restart boinc to clear it and sometimes i have to actually power the machine down to clear it.. i think it has a memory issue because i keep getting xvrm or some such in the logs always pointing to it and to the same locations every time.
the gpl gpumemtest i downloaded also locks up on it at the same spot. it is clearly defective. however it seems to behave nicely if it doesnt have long angles to deal with. i at first thought it was a messed up thermal pad on one of the ram chips so i replaced everything  on the chips with some VERY expensive fiberglass web pads and some shin-etsu thermal compound for the gpu and it still misbehaves. its a first design issue engineering model so i am not afraid to retire it or put it into a machine that i dont care much about.