RHEL3 support? Would like to use RHEL3 with CUDA

Greetings, we’d like to use Red Hat Enterprise 3 as our CUDA platform.

This was the only platform supported during pre-beta, as I recall, but

the problem now is that building the SDK project examples requires

libraries that are not available on RHEL3. Suggestions? Workarounds?

(Report from our sysadmin below.)

JDO

============

The Display Driver, CUDA Toolkit, and CUDA SDK installations succeed on

our RHEL3 lab workstations, building the SDK project examples however

fail as follows:

bash-2.05b$ export PATH=${PATH}:/opt/cuda/bin

bash-2.05b$ export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/opt/cuda/lib

make[1]: Entering directory `/tmp/tmetz/NVIDIA_CUDA_SDK/common’

ar qv ./…/lib/libcutil.a ./…/bin/obj/release/bank_checker.cpp_o

./…/bin/obj/release/cmd_arg_reader.cpp_o

<…>

nvcc -o …/…/bin/obj/release/bitonic.cu_o -c bitonic.cu -Xcompiler

“-m32” -Xcompiler “” -I. -I/opt/cuda/include -I…/…/common/inc -DUNIX -O3

nvcc: error while loading shared libraries: libstdc++.so.6: cannot open

shared object file: No such file or directory

make[1]: *** […/…/bin/obj/release/bitonic.cu_o] Error 127

make[1]: Leaving directory `/tmp/tmetz/NVIDIA_CUDA_SDK/projects/bitonic’

make: *** [projects/bitonic/Makefile.ph_build] Error 2

bash-2.05b$

It appears to require libstdc++.so.6 / gcc 3.4.5, which are not

available on RHEL3.

Cuda is officially supported on RHEL4 32bit only.

The beta build is using the default libraries/gcc available on that distribution.

You can try to add the new libraries to RHEL3 (not easy) or do an upgrade to RHEL4 ( probably easier and cleaner).

With RHEL5 just out, I doubt RHEL3 will be supported in the future, so now it may be a good time to upgrade your distro.

Massimiliano

Unfortunately I am using CUDA for teaching in a university environment with hundreds of machines in the department all running the same distro. For us, that’s RHEL3, because RHEL4 has many fewer instructional (CAD) packages. It is not likely they would upgrade the entire department simply to handle just CUDA when it would break so many other packages.

Do you have any tips for adding the new libraries into RHEL3?

Was CUDA not originally supported on RHEL3 before the beta (Oct-Nov, when we first got it)? That was my understanding - the pre-beta was only on RHEL3, so we did not feel any need to move distributions at that time.

JDO

I think it has always been RHEL4 update 3, maybe that’s why you’re thinking of the 3? Can you just compile the version of gcc/libraries you need?

The first non-internal CUDA build was (I believe) RHEL3 update 5. We had update 8 at the time, but it worked OK. I am confident that it was RHEL3 that was the first supported version.

We’d prefer to do as few custom builds as possible and instead use Red Hat’s approved up2date software distribution system. That’s how we distribute software to our workstations. There are a lot of issues faced by enterprise-class hundreds-of-machines environments that aren’t faced by single-machine installs. Anything that has to be done “by hand” - whether it be “just compile this other software version by hand” or “just rebuild the NVIDIA driver into the kernel by hand every time there’s a kernel update” - is likely not an option for environments like ours. It would be more convenient if CUDA was supported on RHEL3 and worked with the RHEL3 sw tools. I know NVIDIA is committed to supporting lots of Linux distros in the long run; it’s a manpower issue and they’re doing the best they can. Just a vote for RHEL3 being one of those distros. :)

JDO

yes, John, you are correct. The first alpha was RHEL3 update 4 or 5. Based on feedback from our main initial customers, we made the decision to focus on RHEL4 update 3 for the beta release.

Mark

We have other software from vendors that is based on RHEL3. They are taking a long time to get up to Rhel4, likely sometime next year. Before we can deploy that to our user community, we need to test. That means that we won’t be able to upgrade our machines to RHEL4 until late 2008 or early 2009. And yes, we are not happy with them.

So in the meantime, without support for Cuda on Rhel3, we won’t be able to deploy any Cuda based software until then, at least on these machines.

My vote it to support Rhel3.