Choice of Linux Distro Can't get verification binaries to run in SuSE nor Fedora

I have a 9500 GT and a Tesla C1060. New system, 9550 based system. Dual boot with Win XP and whatever Linux I want to load. Got through the installation verification under XP (which I thought would be the tough part) using 32 bit. I loaded SuSE 11.1 and dispite updates and so forth, got to the point of using ‘make’ to create binaries to run…but never got them created. Tried a few times and failed each time. As these are brand new installations, and the last time I did all updates to SuSE before loading the nVidia software that I downloaded. The system seemed to be missing linkage to some key files along the way. So I redid the whole with Fedora 10. Similar problem in that I couldn’t get the make file to create the binaries. This time among the error messages, it indicated that /usr/bin/ld: cannot find -lglut and that collect 2: ld returned 1 excit status.

I wasn’t expecting this level of difficulty to run through the examples to make sure the system is running correctly. Slince I am working with a new computer, I would like to get this working with minimal difficulty. I am not wedded to any particular linux. Would any of the other linux choices work with minimal (or even no) problem? Obviously “free is nice”…I would try Ubuntu next…but I could even go for RedHat enterprise if that runs without a hitch. Reading other posts, sometime you would says something like load libglu-dev or somesuch…where do I find that?

For now, the main question is: Which Linux is most likely to give me the most error-free implementation of the Cuda structure so I can begin working on my project? The second question is: when I was running the SP examples, I noticed they all used the 9500GT…can I direct the examples to use the Tesla card? Is there a better document to read other than the NVIDIA CUDA developmet Tools 2.3 - Installation and Verification on Linux (as well as the XP counterpart)?

P.S. In both the SuSE and Fedora attempts, I noticed that in loading the SDK.run program, I got the message “Could not load CUDA. Enter the Path to CUDA” which I did (I assume it is /usr/local) but since I accepted all defaults, I was surprised it didn’t know the path since it should have know where it put it in the prior step…
P.P.S I only want to use a 64 bit Linux. I forgot to mention that above. Also, my system has 4 GB of memory and I gnerally have it overclocked to about 3.2 GHz. If my programs work out, I will be upgrading to an i7 system, with a faster processor and probably 6 or more GB of memory…and possibly even multiple Tesla cards with more memory again.

Dick

The OpenGL SDK samples require GLUT (OpenGL utilities library), both the library and the development files. I don’t think any distributions install them by default. but it should be trivial to install them from the package manifest of whatever distribution you are using).

It is only the OpenGL samples that need those extra libraries. The “core” verification examples (bandwidthtest, devicequery) don’t need anything else and should build and run without intervention.

Ubuntu 9.04 or Redhat/Centos/SL 5.3 both work fine in my experience. Neither include those GL development headers in the default install, though, but as mentioned above, it is trivial to install them with either apt/aptitude or yum depending on which distro you choose.

Use nvidia-smi to mark the 9500GT as compute prohibited, and the driver will automagically use the Tesla. Many of the examples also accept a gpu number as a command argument.

Not really. The assumption seems to be that whomever is installing the development tool chain possesses some level of experience with software and OS administration (keeping in mind that the driver packages contain the bare necessities to run built CUDA applications, and developers are responsible for delivering anything else as part of their software package). It seems that assumption maybe misplaced in some cases, as a lot of students and first time users seem to be installing Linux specifically to work with CUDA. I am not aware of any documentation that is more accessible than what is provided, I am afraid.

which latest distro of linux can we use for cuda 3?

The download page:

http://forums.nvidia.com/index.php?showtopic=149959

has CUDA toolkit packages for all the supported distributions. That list should answer your question.

As a small addition to the list of officially supported distributions, I can confirm that CUDA packages (from early released, up to and including 3.0 Beta) work fine on Slackware, that is in turn well known as “vanilla” distribution, so I guess it is safe to state that CUDA will work well with any modern Linux distribution; of course, certain level of knowledge in working with Linux system is needed for installing CUDA, and then compiling and running CUDA codes…

For Fedora, CentOS, RHEL, etc., you need to do the following:

yum install freeglut-devel

yum install libXi-devel

yum install libXmu-devel

before making the SDK applications. It’s probably the same with the other distros.

Note that for Fedora 11 or Fedora 12, you will have to add the following line in C/common/common.mk

–compiler-options -fno-inline

to NVCCFLAGS

http://forums.nvidia.com/index.php?showtopic=104525

The usual source of compatibility problems with CUDA are new versions of gcc. The CUDA toolkit often works with distributions newer than the approved and tested list, except when there is a new version of gcc, then things break.