How to get CUDA work (GNU/Linux)
24 Apr 2011 19:05:10 UTC
Topic 195765
(moderation:
Hi all !
I'm running einstein@home on Boinc 6.10.58 on a GNU/Linux Fedora 14.
My Graphic card is CUDA-capable (GeForce GT240) but it seems that i don't receive CUDA work
I'm using the nvidia proprietary driver 260.19
Is there a way to get CUDA work in this platform?
Cheers
GNU/Linux <3
How to get CUDA work (GNU/Linux)
)
Your machine got plenty of CUDA tasks, but returned all errors, like this.
Not fully sure what is wrong there, I've seen this error (999) once when the driver didn't match the kernel module (a reboot should fix this).
BM
BM
Yes, that's it, exactly !
)
Yes, that's it, exactly !
It's strange because my graphic card work as usual but i'll try to reboot and request a task with CUDA
GNU/Linux <3
Just a simple reboot and it's
)
Just a simple reboot and it's ok
Thank you !
GNU/Linux <3
I know, never change a
)
I know, never change a running system...but NVIDIA has released official 270.* drivers for Linux.
Note that for Linux, this actually makes a difference.
If you have 260.* drivers installed, a BRP3 CUDA task running on your Linux PC will take an entire CPU core in addition to the GPU.
If you have a 270.* driver installed, a BRP3 CUDA task will use less than a pull CPU core . So on a PC with N CPU cores and 1 GPU, under the 260 driver you'll see up to N tasks running in parallel, while with the 270 driver you can get N+1 jobs running in total.
HB
RE: I know, never change a
)
but then of course you'll see a performance decrease. it's faster to free a core for the GPU-task. either with ncpus= x-1 in cc_config.xml or via the processor-usage settings in boinc-manager.
RE: but then of course
)
The CUDA task will run slower, yes, but the overall productivity should be higher because you get another CPU job running.
CU
HB
@Bikeman : It's very
)
@Bikeman :
It's very interresting, sad that i'm a bit afraid about installing nvidia .run driver from the website instead of the repo driver (which begin to be quite old, i know) it's also my main desktop, i don't want to broke it too often :p
Fedora 15's rpm fusion repo provide the nvidia 270.x series, so i'll surely switch to it after the beta stage of Fedora 15.
I'm interrested about one thing :
Does this mean that if i'm running a CUDA task (always running now), it's better to set 75% to
On multiprocessors systems, use at most ... of the processors?I'm running boinc on a dual core hyperthreading (Intel Core i3)
GNU/Linux <3
RE: Does this mean that if
)
i can confirm that for my GTX260 and a mates GTX470.
you'd probably gain some extra % of total output on fast cards when running an app_info with 2 or even 3 tasks at a time and enough free CPU power.
but since the current app depends a lot on CPU speed, you need to figure out what works best on your setup.