9

Can someone suggest the best linux distro for CUDA development. The reason why I'm asking is that I tried installing the latest cuda SDK in Fedora 12, and it was a real pain in the neck. It took me 8 hours to get the nouveau driver removed and the nvidia-driver installed. After that somehow the OS decides to act up and blow up the /var/log/message file to 9 GB and eat up all my remaining space, with strange errors. I don't even understand what happened further, but my Nvidia drives don't work anymore. Please don't flame me, I'm NOT a windows fanboy or anything. I've been using Linux since 2002, and actually like it. Its just my personal experience. Would be really helpful for positive suggestions. Fanboys, please stay aside.

Thanks in advance.

Improve this question
3
  • Problem with asking for a "best x for y" is that every person will have their own opinion, which would end up confusing you.
    – Sathyajith Bhat ♦
    Commented May 27, 2010 at 17:15
  • When I meant best, I actually meant based on the amount of effort required to get going. I'm sufficiently proficient in linux shell. However, I don't do much of driver installing uninstalling stuff..
    – 0fnt
    Commented May 27, 2010 at 17:51
  • Not going to answer, as it's anecdotal, but I had the same experience getting the nvidia drivers installed on Ubuntu 10.04. Needed to blacklist a few nouveau and framebuffer modules, and needed to specify the kernel version in order to get the nvidia drivers installed. Once installed though, CUDA installation is quite straightforward and works like a charm.
    – fideli
    Commented May 28, 2010 at 4:02

5 Answers 5

Reset to default
5

I would use Arch Linux since it comes with the very latest stuff and you can install CUDA-SDK/everything you need from AUR directly. (Also NVidia driver is directly avaliable in the repository.) But yeah you can use anything if you build up there the environment.

Maybe Fedora also fits the job with the very latest stuff.

Improve this answer
2

i used OpenSuse and it worked pretty smoothly and was easy to set up. extra benefit is OpenSuse is a pretty nice desktop environment w/nice package management tools that doesn't suffer from limited functionality such as ubuntu's over-stripped down-ness (and brown? really?).

on windows now, though, so i can use it for gaming when i'm not developing. but i can say for cuda development opensuse worked great.

Improve this answer
1

I tend to think that the best distro for this sort of thing is that which the developers of the thing use: any instructions and downloads are far more likely to work on the distro they were written for than on any other.

Looking at the CUDA website I would try getting hold of one of those distros - including the version specified. Of course, when you want to install a newer version of something you get into a whole new world of pain...

Improve this answer
1

Disclaimer: You may not like this answer...

I have a dual-boot system. I use KUbuntu-10.04 amd64 as my primary operating system. I wish it had gcc 4.5 because I want OpenMP 3.0, but otherwise... I've installed CUDA 3.1, CUDA 3.2, and the latest drivers (260.24) up and running. I haven't had any issues installing drivers or CUDA. I also use nvidia-smi and the following script called "cuda" which I've placed in /etc/init.d/ and start on all runlevels:

#!/bin/bash

/sbin/modprobe nvidia

if [ "$?" -eq 0 ]; then

# Count the number of NVIDIA controllers found.
N3D=`lspci | grep -i NVIDIA | grep "3D controller" | wc -l`
NVGA=`lspci | grep -i NVIDIA | grep "VGA compatible controller" | wc -l`

N=`expr $N3D + $NVGA - 1`
for i in `seq 0 $N`; do
mknod -m 666 /dev/nvidia$i c 195 $i;
done

mknod -m 666 /dev/nvidiactl c 195 255

else
exit 1
fi

nvidia-smi --loop-continuously --interval=60 --filename=/var/log/nvidia-smi.log &
nvidia-smi -g 0 -c 2
nvidia-smi -g 1 -c 1
nvidia-smi -g 2 -c 1

Note: I have three cards, one dedicated to display and two dedicated to CUDA computation.

Now that I've said that, at present, the only minimally good CUDA profiling that I've found is nSight which only works on Windows 7 or Vista with either VS 2005 or 2008 -- they claim it works on VS 2010... riiight. I've been asking about it here. It appears from your other questions that you have access to VS 2010 on Win 7. I'd recommend working in VS 2008 (if possible) for now with nSight 1.5 until nVidia gets the bugs out of the newer build customizations shtuff. If you're either a student or a small business user, you can get a free copy of VS 2005 or 2008 from DreamSpark or BizSpark respectively.

I've also asked about visualizing Win 7 so I don't have to keep rebooting -- but I can only post one link (as I don't have 10 points yet), so I'll put it in a comment. I also have no experience yet with running CUDA under a visualized OS.

ps. If, like me, your doing parallel computing in both CUDA and OpenMP, you might consider any of the Ubuntu 10.10 flavors which were released (RC) yesterday and will be pushed stable in a week. Ubuntu 10.10 has the option of integrating gcc 4.5.

Improve this answer
1

I like Scientific Linux, based on RHEL, for the scientific base lining. For new users, to avoid the nuisance with nouveau, I'd stay with the 5.x series. Otherwise, for 6.x, be sure to install the ElRepo repository and blacklist nouveau. When you update, the driver will automatically be recompiled.

Improve this answer
1
  • Why is ElRepo useful for CUDA on RHEL 6.x?
    – Christian Hudon
    Commented Aug 15, 2014 at 15:13

Not the answer you're looking for? Browse other questions tagged .