[gmx-users] Cuda CC 2.0 restrictions

[Szilárd Páll]

On Tue, Sep 9, 2014 at 10:58 PM, Mirco Wahab
<mirco.wahab at chemie.tu-freiberg.de> wrote:
> Thank you very much for your answer. After reading
> Mark's very detailed posting on the topic, I played
> with some parameters but did not manage to get under
> 64K grid dimension.
>
> On 09.09.2014 21:35, Szilárd Páll wrote:
>>
>> Is this rather large box a system that can actually be simulated with
>> a useful speed on a single Fermi GPU? Even with 5 fs time-step you
>> won't get much more than 1-1.5 ns/day on a fast Fermi GPU like a GTX
>> 580.
>
>
> This is a rather simple toy system used for machine comparison and
> as a hardware reliability test. ~36nm^3 SPC water box, ~1.56x10^6
> water (GTX-780 on 3770K is 0.45ns/day).
>
> I posted the question because I did not understand why this
> happened on the GTX-580. mdrun -nstlist x did not work (4.6.7).

Indeed, -nstlist got exposed as a commnad line option in 5.0. You can
use the GMX_NSTLIST environment variable instead.

BTW regarding benchmarking: with small systems and single-GPU runs
make sure to disable GPU timing (set GMX_DISABLE_CUDA_TIMING)!

>> Given that you are quite a bit above the limit, unless you are using a
>> quite large nstlist, you may not be able to decrease it enough to fit
>> the system on the GPU. What you can do instead is to use
>> domain-decomposition and start multiple ranks per GPU. In your case
>> two-way DD should be enough.
>
>
> This sounds interesting. I'll try to accomplish that (I never thought
> this *this* is possible ...).

Not just possible, but it is in fact advantageous from performance
point of view with many cores/GPU!

> Thank you very much (and Mark)
>
> Mirco
>
>
>
>
> --
> Gromacs Users mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a
> mail to gmx-users-request at gromacs.org.