[gmx-users] Workstation choice

[Szilárd Páll]

BTW, I'd recommend caution when using the dated d.dppc benchmark for
drawing performance conclusions both because it may not be too
representative of other workloads (small size, peculiar settings) and
because it uses all-bonds constrained with 2fs time step which is not
recommended these days (unless using larger time steps). Constraining only
h-bonds is generally considered best practice and it is faster, especially
with GPUs.

I'd also like to warn about drawing conclusions about hypothetical
production setups from performance measurements with under-utilized CPUs.
Leaving cores empty creates conditions which are not representative to
real-world scenarios (unless of course you plan to leave those core empty
during production runs too). The least I'd do is to put some CPU load on
those cores (e.g. on an 8-core CPU run mdrun on 4 cores + 1 GPU and keep
the other 4 cores busy with some dummy load like a stress tester tool or a
CPU-only mdrun run).

--
Szilárd


On Tue, Sep 11, 2018 at 11:13 PM Wahab Mirco <
Mirco.Wahab at chemie.tu-freiberg.de> wrote:

> On 11.09.2018 22:37, Olga Selyutina wrote:
> >> How would the 8600K perform without GPU (using the "-nb cpu"
> >> option on the mdrun binary)?
> > 8600K  with frequency 4600Mhz was used
> > Command line:
> >   gmx mdrun -nb cpu
> >
> >   Core t (s)   Wall t (s)        (%)
> >        Time:     1529.942      254.990      600.0
> >                  (ns/day)    (hour/ns)
> > Performance:      169.419        0.142
>
> Interesting - That means, for your record:
>
>    i5-8600K, 6C/6T, AVX2_512: 169.419 ns/day
>    R2-2700X, 8C/16T, AVX2_128: 152.728 ns/day
>    R1-1700X, 8C/16T, AVX2_128: 141.456 ns/day
>
> your overclocked coffee-lake i5 would already
> be slightly faster than a Ryzen2 2700X.
>
> But if you want to simulate larger membrane systems,
> you should possibly test also on larger membrane
> systems. There's a DPPC/Water membrane benchmark included
> in the Gromacs benchmark set. Unfortunately, it (d.dppc)
> wouldn't work out of the box with newer versions.
> I patched it up and converted it into a NPT membrane-benchmark:
> http://gromacs.deppenfraktur.de/gmxbench-d.dppc-npt.tgz
> and already got some results for some of our antique boxes:
>
> Results for 2018.3:
>
>    ***                     i5-7600K, 4C/4T, AVX2_512: 5.512 ns/day
>    ****                    i7-6700K, 4C/8T, AVX2_256: 6.858 ns/day
>    ****                    R1-1700X, 8C/16T, AVX2_128: 7.490 ns/day
>    *****                   i7-5820K, 6C/12T, AVX2_256: 9.162 ns/day
>    *****                   R-2 2700X, 8C/16T, AVX2_128: 9.401 ns/day
>
>    **************          GTX-980, i5-7600K, 4C/4T, AVX2_512: 28.071
>    **********************  GTX-1080, R1-1700X, 8C/16T, AVX2_128: 42.908
>
> If you get some people with contemporary hardware to test for
> such workloads related to your projects, you can possibly buy
> boxes accordingly.
>
>
> M.
> --
> 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.
>