Just got a new machine, spent some time installing newest OpenSuSE and configuring. Got to the point of installing NVidia CUDA SDK (I’ll be using this machine for CUDA development, so it’s a necessity), and discovered that it won’t install in 12.2, only 12.1. Is there an easier way to downgrade the version than wiping the disk and starting from scratch?
Thanks,
James
On 2012-11-17 19:06, jamesqf wrote:
>
> Just got a new machine, spent some time installing newest OpenSuSE and
> configuring. Got to the point of installing NVidia CUDA SDK (I’ll be
> using this machine for CUDA development, so it’s a necessity), and
> discovered that it won’t install in 12.2, only 12.1. Is there an easier
> way to downgrade the version than wiping the disk and starting from
> scratch?
a zypper dup, but it is not guaranteed. Not recommended.
–
Cheers / Saludos,
Carlos E. R.
(from 11.4 x86_64 “Celadon” (Minas Tirith))
On 11/17/2012 07:06 PM, jamesqf wrote:
> it won’t install in 12.2, only 12.1
why?
error message?
anyway, 12.2 is about a month old, it could be that nVidia will catch
up, soon? might ask’em…
i’d suggest the best move from 12.2 to 12.1 would be format and install…
It complains about an unsupported version of gcc. 12.2 has gcc 4.7.1, the CUDA installer wants gcc 4.3.1. And of course it quits there, so who knows what further problems it might encounter if it gets past that.
anyway, 12.2 is about a month old, it could be that nVidia will catch
up, soon? might ask’em…
I don’t know. NVidia doesn’t seem big on support - it’s more like “You will use what we tell you to use, foolish mortal!” With a 0.4 difference in supported version already, I can’t see them getting up to date anytime soon.
i’d suggest the best move from 12.2 to 12.1 would be format and install…
Not quite what I wanted to hear, but if I must…
On 11/17/2012 10:56 PM, jamesqf wrote:
> Not quite what I wanted to hear, but if I must…
understand your point…
but that is the penalty for installing 12.2 when the CUDA page
https://developer.nvidia.com/cuda-downloads specifies 12.1 is
supported, and doesn’t mention 12.2! 
sure, you could try a reverse upgrade (with zypper dup) but i would be
astonished if you found that troublefree!!
here are the suggested/supported upgrade paths (12.1 to 12.2)
http://tinyurl.com/35p966c
http://tinyurl.com/93uemsr
http://tinyurl.com/7l4m2td
each of those are known to not be trouble free–note the requirements
and BACKUP suggestions…
couple that with the fact that there is no recommended/supported
downgrade path, and then you know why format/install has been
recommended…
ON THE OTHER HAND: it seems you are gonna do development (earn money
with this machine, no?) if so you might be well advised to go ahead and
install the latest Enterprise version…if for no other reason, because
openSUSE 12.1 will go past its end of life on May 15th 2013…
after which you will be running a system with no easy way to keep it
secure (because there will be no more security patches after end of life)
and, that short life span is a problem you will not run into with any of
these (using CUDA page nomenclature):
SUSE Server 11 SP1
SUSE Server 11 SP2
RHEL 5.X
RHEL 6.X
well, i’m not certain when any of those drop dead, but i would be
shocked to learn any of them go unsupported in six months–and, what
will you do when openSUSE 12.1 goes unspported?
hmmmm…i also don’t know the drop dead dates of the CUDA supported
fedora or ubuntu offerings…
Am 17.11.2012 22:56, schrieb jamesqf:
> It complains about an unsupported version of gcc. 12.2 has gcc 4.7.1,
> the CUDA installer wants gcc 4.3.1. And of course it quits there, so
> who knows what further problems it might encounter if it gets past that.
>
As it was always possible to have several versions of gcc at the same
time on an openSUSE system you could try to install gcc43 from the
devel:gcc repository. Depending on what the CUDA SDK documentation says
you will have to check which environment variables to set so that it
uses the gcc43.
I had myself in the past several gcc versions in parallel without
problems, not sure how good that works with 12.2.
–
PC: oS 12.2 x86_64 | i7-2600@3.40GHz | 16GB | KDE 4.8.5 | GeForce GT 420
ThinkPad E320: oS 12.2 x86_64 | i3@2.30GHz | 8GB | KDE 4.9.3 | HD 3000
eCAFE 800: oS 11.4 i586 | AMD Geode LX 800@500MHz | 512MB | lamp server
Now how many competently-written packages in the world have specific OS version requirements? You install the latest OS, then add such software as you need, no?
after which you will be running a system with no easy way to keep it
secure (because there will be no more security patches after end of life)
Not an issue.
Am 18.11.2012 18:46, schrieb jamesqf:
> Now how many competently-written packages in the world have specific OS
> version requirements? You install the latest OS, then add such software
> as you need, no?
No as your claim is naive, the CUDA SDK contains a compiler which
heavily depends on having a compatible gcc compiler on your system to
work as it builds on top of its capabilities.
Later gcc versions introduced so many incompatibilities to previous
versions that the cuda compiler does not work with them.
You may want to look how fedora people make cuda run on newer fedora
versions
http://fedoraproject.org/wiki/Cuda
scroll down to fedora 17 and you see it works for them with a
compatibility package (in openSUSE this would be gcc33 instead of
gcc34), I have not tested this but you could give it a try or you use as
I mentioned before the gcc43 from the devel:gcc repository.
–
PC: oS 12.2 x86_64 | i7-2600@3.40GHz | 16GB | KDE 4.8.5 | GeForce GT 420
ThinkPad E320: oS 12.2 x86_64 | i3@2.30GHz | 8GB | KDE 4.9.3 | HD 3000
eCAFE 800: oS 11.4 i586 | AMD Geode LX 800@500MHz | 512MB | lamp server
Not really that naive, since I’ve been using CUDA for several years (about as long as it’s been generally available), and have not had problems installing whatever version of the CUDA installer was current on whichever version of openSuSE I had on the machine - which from memory are 10.x and 11.x versions.
You may want to look how fedora people make cuda run on newer fedora
versions…
At this point I’d be interested in anything that will work. 12.1 install hangs from both net and DVD .iso (always after installing the ~144 MByte base system package), Fedora installs but won’t run X, Ubuntu tries to be Windoze… Honestly, I’ve been using Linux since the mid-90s (mostly SuSE), and have never had problems anything like this before. Usually it just installs and runs.
What was the point yet ?
Was it: “Is there an easier way to downgrade a 12.2 install to 12.1 than to do a fresh install” ?
For sure not.
Did you ever hear about a way to downgrade a Windows 8 install to a Windows 7 install by an upgrade ??
I mean:
The most important things in life are simple.
So do a fresh install, and be sure of what you got !
Best wishes
Mike
If you can wait a few days I will give CUDA 5 a try on my 12.2. I have
to wait till my new graphics card arrives. At the moment I run my PC in
contrast my signature which has a nvidia in it with a cheap Radeon card
till I have my new card (GTX 650 Ti).
If you did not already find a way to make it run on 12.2 and if I can
make it run I will post the steps to this thread.
PS: I did not mean that you are naive but that the way you stated the
independence of a software from the underlying operating system is in
that generality a bit naive. I did not intend to offend you (English is
not my native language).
–
PC: oS 12.2 x86_64 | i7-2600@3.40GHz | 16GB | KDE 4.8.5 | GeForce GT 420
ThinkPad E320: oS 12.2 x86_64 | i3@2.30GHz | 8GB | KDE 4.9.3 | HD 3000
eCAFE 800: oS 11.4 i586 | AMD Geode LX 800@500MHz | 512MB | lamp server
Am 18.11.2012 22:28, schrieb Martin Helm:
> If you can wait a few days I will give CUDA 5 a try on my 12.2. I have
> to wait till my new graphics card arrives. At the moment I run my PC in
> contrast my signature which has a nvidia in it with a cheap Radeon card
> till I have my new card (GTX 650 Ti).
> If you did not already find a way to make it run on 12.2 and if I can
> make it run I will post the steps to this thread.
@jamesqf
There is no need for you to downgrade to 12.1, CUDA 5 works with 12.2. I
did a quick test here on my Dell Precission with nvidia card.
First add the repository
http://download.opensuse.org/repositories/devel:/gcc/openSUSE_12.2
and install gcc43 and gcc43-c++
Then you run the installer with the -override option to skip all
compatibility tests. I installed to the default /usr/local/cuda directory.
Create a symlink /usr/bin/gcc-4.3 to /usr/local/bin/cuda
Adjust your LD_LIBRARY_PATH to have /usr/local/cuda/lib64 and the PATH
to contain /usr/local/cuda/bin
Now you can compile for example fluidGL and it works.
–
PC: oS 12.2 x86_64 | i7-2600@3.40GHz | 16GB | KDE 4.8.5 | GeForce GT 420
ThinkPad E320: oS 12.2 x86_64 | i3@2.30GHz | 8GB | KDE 4.9.3 | HD 3000
eCAFE 800: oS 11.4 i586 | AMD Geode LX 800@500MHz | 512MB | lamp server
Am 18.11.2012 23:49, schrieb Martin Helm:
> Create a symlink /usr/bin/gcc-4.3 to /usr/local/bin/cuda
That shall of course be
ln -s /usr/bin/gcc-4.3 /usr/local/cuda/bin/gcc
sorry for the fuzzy description and typo.
–
PC: oS 12.2 x86_64 | i7-2600@3.40GHz | 16GB | KDE 4.8.5 | GeForce GT 420
ThinkPad E320: oS 12.2 x86_64 | i3@2.30GHz | 8GB | KDE 4.9.3 | HD 3000
eCAFE 800: oS 11.4 i586 | AMD Geode LX 800@500MHz | 512MB | lamp server
Not offended at all, just trying to be clear about things - not the easiest task, even though English is my native language. If it appears otherwise, put it down to my lach of writing ability. And I am grateful for the help.
Though I must say that I manage to write fairly complex software that runs on any flavor of *nix, and with a few #ifdefs to handle file system differences, on Windows as well. It’s only in the last few years that I’ve increasingly run into things that require e.g. having particular versions of dlls.
No, but then I know very little about Windows - and anyway, isn’t Linux supposed to be BETTER than Windoze?
In any case, thanks for the suggestions, everyone. I’m back re-installing 12.2 (12.1 won’t install, Fedora won;t run right) , and when that’s done will try installing the older gcc,
On 2012-11-19 00:56, jamesqf wrote:
> No, but then I know very little about Windows - and anyway, isn’t Linux
> supposed to be BETTER than Windoze?
As said, the only known way is a zypper dup. However, nobody tests this
procedure, you will be making your own path.
There are tales of people that did it and survived to tell, though.
–
Cheers / Saludos,
Carlos E. R.
(from 11.4 x86_64 “Celadon” (Minas Tirith))
That was close. I’m not sure how to get the installer (I use Yast2) to override, so wound up with both gcc versions in /usr/bin. Then just changed the symlink “gcc” to point to 4.3 instead of 4.7, and that got me started. Needed to install a few more packages (mostly related to glut) to get the samples installed, but I now get a working sample to compile. Whew!
Thanks, everyone!
PS: If this is so basically simple, I wonder why the NVidia people don’t mention it?
Am 19.11.2012 04:56, schrieb jamesqf:
> override, so wound up with both gcc versions in /usr/bin. Then just
> changed the symlink “gcc” to point to 4.3 instead of 4.7, and that got
> me started.
A word of caution, if you ever need to compile a kernel module change
the link back to 4.7 before (the system relies in that case on having
the right compiler which was used for the kernel after that you can
change it again to 4.3). Don’t know if you ever need to do that but keep
it in mind.
The -override option was for the cuda installer not gcc
sh cuda_5.0.35_linux_64_suse12.1-1.run -override
sorry for not being explicit, it was late night for me yesterday.
> PS: If this is so basically simple, I wonder why the NVidia people
> don’t mention it?
>
Because they do not care and rely that people find that out on their own?
Honestly I would have expected something much simpler like a command
line switch in the installer to tell the paths of a custom compiler or
something similar.
Glad to read it works for you now, have much fun.
–
PC: oS 12.2 x86_64 | i7-2600@3.40GHz | 16GB | KDE 4.8.5 | GeForce GT 420
ThinkPad E320: oS 12.2 x86_64 | i3@2.30GHz | 8GB | KDE 4.9.3 | HD 3000
eCAFE 800: oS 11.4 i586 | AMD Geode LX 800@500MHz | 512MB | lamp server
Sure. I figure I’ll write a script (like I have for openMPI) that sets PATH,
LD_LIBRARY_PATH, and correct compiler links.
> PS: If this is so basically simple, I wonder why the NVidia people
> don’t mention it?
>
Because they do not care and rely that people find that out on their own?
Honestly I would have expected something much simpler like a command
line switch in the installer to tell the paths of a custom compiler or
something similar.
So once again, we have to reinvent the wheel
Now that’s all a bit funny.
Clearly not.
Why do you run a Linux system at all?
Hmm, I wrote a finite element programme on a Mac and some data aquisition code using interrupts for DOS.
I would conclude that you know a bit about differences and changes on different systems and should be
able to cope with them.
Really?
If 12.1 won’t install for you, then the guys & dolls here will clearly try to help you out.
Mike