Re: running CUDA cards
On Tue, May 31, 2011 at 5:56 PM, Lennart Sorensen
<lsorense@csclub.uwaterloo.ca> wrote:
> On Tue, May 31, 2011 at 05:20:35PM +0200, Francesco Pietra wrote:
>> I have just set up a gaming machine with
>>
>> Gigabyte GA 890FXUD5
>> AMD Phenom II 10775T
>> 2 x GTX470 GPU cards
>> 4 x 4GB RAM
>> 2 x 1 Tb HD for RAID1
>>
>> and need to install amd64 to run molecular dynamics using (free for
>> non-commercial use) NAMD software (released binary below or
>> compilation from source). All that is experimental, with little
>> experience and I have no experience whatsoever with CUDA cards. My
>> question is about the version of amd6a to be best used (lenny or
>> squeeze) and what should be added to the typical server installation
>> according to the requirements:
>
> Lenny will stop having support soon, so absolutely go with squeeze.
> The nvidia-glx package in squeeze supports the GTX470 card. Lenny does
> not. So install squeeze.
>
> To install the driver, add 'contrib non-free' to your lines in
> /etc/apt/sources.list then do:
>
> apt-get update
> apt-get install nvidia-kernel-dkms
> apt-get install nvidia-glx nvidia-glx-dev libcuda1-dev libcuda1
>
>> (1) NVIDIA Linux driver version 195.17 or newer (released Linux
>> binaries are built with CUDA 2.3, but can be built with newer versions
>> as well).
>
> squeeze has 195.36.31 so that should work.
>
>> (2) libcudart.so.2 included with the binary (the one copied from the
>> version of CUDA it was built with) must be in a directory in your
>> LD_LIBRARY_PATH before any other libcudart.so libraries. For example:
>>
>> setenv LD_LIBRARY_PATH ".:$LD_LIBRARY_PATH"
>> (or LD_LIBRARY_PATH=".:$LD_LIBRARY_PATH"; export LD_LIBRARY_PATH)
>> ./namd2 +idlepoll <configfile>
>> ./charmrun ++local +p4 ./namd2 +idlepoll <configfile>
>
> The libcuda1 package should probably take care of the library I think,
> but maybe not.
Unlike CPU machines, where the computational software is "stable",
with CUDA support the situation is highly experimental, changes
rapidly and has requirements of recent OS support. New versions of
NAMD every few weeks and even night builds. One of the goals is to
have even energies computed by GPU. That to say that I would have no
objection to install wheezy unless it is too much an adventure. It
might well be that a new version of NAMD comes soon with requirements
of libraries not available in squeeze. Finally, NAMD would be the only
application installed on this dedicated machine so that if dramatic
problems occur, there will be little to reinstall. I imagine that
amd64 RAID1 support will be at least as good as with lenny. In this
case, perhaps, hardware RAID1 would be better but it is too expensive.
I plan to change from lenny to squeeze with the CPU workstation.
Quantum mechanics unlike classical mechanics has no CUDA support yet
(except with very expensive proprietary software, if it works at all).
Thanks a lot
francesco pietra
>
> wheezy has a lot more cuda packages available, and much newer drivers too,
> but is of course testing, not stable.
>
>> THE FOLLOWING CAN BE SKIPPED, unless one is specifically interested in
>> the matter: The +idlepoll in the command line is needed to poll the
>> GPU for results rather than sleeping while idle, i.e. NAMD does not
>> use any non-specified GPU card. Each namd2 process can use only one
>> GPU. Therefore you will need to run at least one process for each GPU
>> you want to use. Multiple processes can share a single GPU, usually
>> with an increase in performance. NAMD will automatically distribute
>> processes equally among the GPUs on a node. Specific GPU device IDs
>> can be requested via the +devices argument on the namd2 command line,
>> for example:
>> ./charmrun ++local +p4 ./namd2 +idlepoll +devices 0,2 <configfile>
>
> --
> Len Sorensen
>
Reply to: