I am just trying to figure out which linux distribution to use for CUDA development. I have used Fedora 10 but have since decided to abandon it, there are far to many issues with compatability at this time. If you could give me some advice on what is, in your opinion, the best\most stable distribution that would be great. Please if you are going to say “It doesn’t matter which distribution you are going to use, it is just preference” just save your time because I am looking for helpful posts not “you’re a noob comments”.
If you want to make sure it works great with CUDA, the best option is probably redhat enterprise linux 5.x
(or centos 5.x if you want a free (as in beer) version)
The default compiler in CentOS 5.* (and RHEL 5.*) will always be gcc-4.1. RHEL/CentOS-5.3 will have gcc-4.3 available as a technology preview (a gcc43 package which will provide /usr/bin/gcc43).
Even though gcc-4.3 is not supported by nvidia, it’s the default compiler for Fedora 9 and 10. I’ve not had any problems so far with CUDA 2.1 on Fedora 10.
I’m using Ubuntu 8.04 with no problems. I have both an Nvidia graphics card and a Tesla card. However, you have to install the Nvidia graphics card drivers from Nvidia, not from the Ubuntu repository. If you have a problem with low resolution, there is a command you can use to solve this. See my earlier post on this forum. One annoying issue is that if you do an Ubuntu update, it may erase your Nvidia drivers and you may have to install both your graphics driver and your Cuda driver as well as reissue the command to solve the low res problem. I’m guessing that Ubuntu does not like non-open source drivers. I hope that Nvidia will work with Ubuntu to fix this as overall I prefer Ubuntu and this is the only small issue that I have found.
Having abandoned fedora myself back about a year ago, you’ve made a right choice :)
If you’re going to go 8.10, be warned you’ll probably have to downgrade gcc - once you get the error messages from the compiler, google for them and it will bring up the forum post with the solution,
Also, gcc is not installed by default: sudo apt-get install build-essentials will give you the compiler and toolchain. Last time I played with it, apt-get was a LOT faster than yum :)
Note you won’t have to install the drivers from nvidia directly: you can get the 180 driver by explicitly commanding: sudo apt-get install nvidia-glx-180 (or at least it will tell you what you actually need to do it.
All this is from the main repo - no livna required. Plus, it will automatically sort out proprietary codecs that you may need - something that used to be a real hassle in fedora.
Is OpenSUSE 11.1 supported ? The release notes for CUDA 2.1 just say OpenSUSE 11, so CUDA 2.1 presumably pre-dates OpenSUSE 11.1 ? Any known issues, as I’m considering going for 11.1 (currently on 10.2 which is no longer supported) ?
imho Debian is the most stable Linux OS
(Ubunti iss what Debian user call unstable External Image )
EDIT:
since I use cuda, the cuda kernel-module may
freeze the whole system if your cuda application has
little bugs…
so you can use the stablest system and bring it down
with a tainted kernel-module…
