How to get Boinc 6.4.5 to use CUDA

[Kobuck]

I've installed the Boinc 6.4.5 x86_64 using F10 update package from Fedora repo.
The client runs fine using the CPUs, but doesn't fing the nVidia display processors.

I've installed nVidia drivers using Fusion repo packages for F10 ( currently 180.25 )
nVidia display settings seems happy and correctly reports driver 180.25 in place as well as 9600 GSO graphics card and LCD monitor.

My objective is to get Boinc using the CUDA capabilities of my GeForce 9600 GSO graphics card and see some of the performance benefits advertised.
Although Boinc client is running it will only use the CPU to do work. I have not been able to discover any flags or configuration settings to help Boinc find CUDA library files.

I'm stuck at the limits of my knowledge. Any help would be appreciated. ( including a more appropriate forum if required)

The following are the startup messages from Boinc. I'm assuming the fact that it can't find the library causes the inability to see the GPU coprocessor.


....
08-Feb-2009 11:10:42 [---] Starting BOINC client version 6.4.5 for x86_64-pc-linux-gnu
08-Feb-2009 11:10:42 [---] log flags: task, file_xfer, sched_ops
08-Feb-2009 11:10:42 [---] Libraries: libcurl/7.18.2 NSS/3.12.2.0 zlib/1.2.3 libidn/0.6.14 libssh2/0.18
08-Feb-2009 11:10:42 [---] Running as a daemon
08-Feb-2009 11:10:42 [---] Data directory: /var/lib/boinc
08-Feb-2009 11:10:42 [---] Processor: 4 AuthenticAMD AMD Phenom(tm) 9850 Quad-Core Processor [Family 16 Model 2 Stepping 3]
08-Feb-2009 11:10:42 [---] Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm 3dnowext 3dnow constant_tsc rep_good nopl pni monitor cx16 popcnt lahf_lm cmp_legacy svm extapic cr8
08-Feb-2009 11:10:42 [---] OS: Linux: 2.6.27.12-170.2.5.fc10.x86_64
08-Feb-2009 11:10:42 [---] Memory: 3.87 GB physical, 5.81 GB virtual
08-Feb-2009 11:10:42 [---] Disk: 727.47 GB total, 647.15 GB free
08-Feb-2009 11:10:42 [---] Local time is UTC -5 hours
08-Feb-2009 11:10:42 [---] Not using a proxy
08-Feb-2009 11:10:42 [---] Can't load library libcudart
08-Feb-2009 11:10:42 [---] No coprocessors
....

System configuration:
ASUS M3N-HT Deluxe Series
• nForce 780a SLI chipset
• integrated nVidia 8200 GPU ( disabled in BIOS )
• BIOS v 1701 2008/12/12
AMD Phenom 9850 Quad-Core Processor
CORSAIR TWIN2X4096-6400C5
• 4GB DDR2 memory @ 800Mhz
• BIOS configured as dual channel
1 x GEFORCE 9600 GSO 768MB
2 x 320GB HD w SATA-II, 7200RPM, 16MB buffer

Linux 2.6.27.12-170.2.5.fc10.x86_64

[lovenemesis]

The same problem happens on mine Fedora 10 i386.

I tried

yum provides */libcudart

, no result was found...

[newiLuvatar]

have a look at this... http://boinc.berkeley.edu/dev/forum_thread.php?id=3564
i'm also having this issue, but i didn't get to solve it yet...

[drunkahol]

Guys,

libcudart comes from the Cuda Toolkit download from the Nvidia site.

From my initial poking around, I've not found an RPM that provides this yet, but these are early days. I'm not interested in BOINC per se - more coding applications direct for CUDA, but part of our journeys appear to be in parallel.

The installation of the toolkit is pretty trivial and simply involves running the script and editing your PATH and ld.so.conf.

I'm certainly running Cuda code on Fedora 11 x86_64 at the moment, but it is only by downloading and installing the drivers from Nvidia - not by using any of the RPMs that are available.

Hope this helps

Cheers

Duncan

[michael12345]

I had cuda running with the direct from nvidia drivers. 185something and the boinc client a few months ago for a project. As I remember, the project has to support cuda, not just you. Installing the nvidia drivers for linux AFAIK provides you with all the cuda drivers. It's not like Windows where there is the vid drivers and then cuda binaries as well.

[marko]

Originally Posted by drunkahol

Guys,

I'm certainly running Cuda code on Fedora 11 x86_64 at the moment, but it is only by downloading and installing the drivers from Nvidia - not by using any of the RPMs that are available.

Hope this helps

Cheers

Duncan

You can run CUDA apps with the rpmfusion nvidia driver also

[lovenemesis]

Hi, does someone know whether the 256M on board graphic memory is one of the minimum requirements for CUDA ? I have a GeForce 8400M G with only 128M on board graphic memory but I could allocate up to 512M memory from system main memory.

[Kazura]

It depends ... "Many CUDA programs require at least 256 MB of memory attached to the GPU."
from link: http://www.nvidia.com/object/cuda_learn_products.html