[gmx-users] dual gpu setup questions

[Szilárd Páll]

On Sat, Apr 9, 2016 at 12:21 AM, Johannes Wagner
<johannes.wagner at h-its.org> wrote:
> hey all,
> got a few questions for people with dual-gpu setups:
>
> - does anyone have tested and well performing explicit threading parameters for a typical GPU compute node with 2 CPU's + HT and dual GPU? The automatic settings for our hardware (2x8core cpu + HT & dual GPU) take 8MPI x 4OMP, which I find odd. Did not start testing and playing around yet, wanted to ask here before...

That's normal and the choice is the result of extensive testing which
showed that typically narrower ranks (<=6 threads/rank is what the
automation targets) are faster; there are additional benefits to
multiple ranks sharing a GPU like transfer and kernel overlap.

>
> - for anyone using 980ti or Titan X cards, in a clear GPU limited case, whats the maximum GPU utilization one can observe (e.g. using nvidia-smi)?

Single rank 70-80%, with domain decomposition 60-70% is good. You will
get less if the tasks outside the force compute region take time (e.g.
constraints, update, pulling).

> - for anyone using dual 980Ti or Titan X setups, what is the required CPU to have a more or less balanced CPU/GPU utilization?

Depends very much on the input, e.g. with AMBER shorter cutoffs are
used hence there is less short-range work, while CHARMM requires
longer cutoffs and use the slightly slower switch kernels.

The two cards you list are quite fast, so they'll need fast CPUs and
will balance well with 6-8 fast cores (e.g. i7 SBE CPUs) or 10+ lower
clocked cores like those on Xeons. Given that you mention 2x8 Xeon
cores, I can imagine that your runs aren't GPU-bound but are limited
by the CPU.

Cheers,
--
Szilárd

> thanks, Johannes
>
> --
> 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.