[gmx-users] gmx 2019 performance issues

[Szilárd Páll]

On Tue, Jan 15, 2019 at 1:30 PM Tamas Hegedus <tamas at hegelab.org> wrote:

> Hi,
>
> I do not really see an increased performance with gmx 2019 using -bonded
> gpu. I do not see what I miss or misunderstand.
> The only thing I see that all cpu run at ~100% with gmx2018, while some
> of the cpus run only at ~60% with gmx2019.
>

GPUs aren't necessarily faster than CPUs, in particular as GROMACS has fast
CPU code for all compute-intensive algorithms. When your CPU is
sufficiently fast to complete the bonded computation before the GPU is done
with the rest (and that is generally the if there are at least 3-4
cores/GPU), it is normal that you do not see improvement from offloading
that task to the GPU. Consequently, you could even see slight performance
degradation in such cases with -bonded gpu vs -bonded cpu.

One benefit can of course be the lower CPU requirement which means you
might be able to run two runs side-by-side using half of the CPU cores each
using 1-2 GPUs.


> There are: 196382 Atoms
> Speeds comes from 500 ps runs.
>
>  From one of the log files:
> Mapping of GPU IDs to the 4 GPU tasks in the 4 ranks on this node:
>    PP:0,PP:0,PP:2,PME:2
> PP tasks will do (non-perturbed) short-ranged and most bonded
> interactions on the GPU
> PME tasks will do all aspects on the GPU
>
> ------------------------------
> 16 cores 4 GPUs
> gmx 2018 48ns/day
> gmx 2019 54ns/day
>
> gmx mdrun -nt 16 -ntmpi 4 -pin on -v -deffnm md_test -nb gpu -pme gpu
> -npme 1 -gputasks 0123
>
> gmx mdrun -nt 16 -ntmpi 4 -pin on -v -deffnm md_test -nb gpu -bonded gpu
> -pme gpu -npme 1 -gputasks 0123
>
> Since the GPUs are not utilized well (some of them are below 50%), my
> objective is run 2 jobs/node with 8 CPUs and 2 GPUs with higher usage.
>
> ------------------------------
> 8 cores 2 GPUs
> gmx 2018 33 ns/day
> gmx 2019 35 ns/day
>
> gmx mdrun -nt 8 -ntmpi 4 -pin on -v -deffnm md_test -nb gpu -pme gpu
> -npme 1 -gputasks 0033
>
> gmx mdrun -nt 8 -ntmpi 4 -pin on -v -deffnm md_test -nb gpu -bonded gpu
> -pme gpu -npme 1 -gputasks 0022
>
> gmx mdrun -ntomp 2 -ntmpi 4 -pin on -v -deffnm md_test -nb gpu -bonded
> gpu -pme gpu -npme 1 -gputasks 0022
> Changing -nt to -ntomp did not help to increase performance.
>
> And the GPUs are not utilized much better. 1080Ti runs max 60-75%
>
> ------------------------------
> The main question:
> * I use 16 core AMD 2950X with 4 high end GPUs (1080Ti, 2080Ti).
> * GPUs does not run at 100%, so I would like load more on them and
> possibly run 2 gmx jobs on the same node.
>
> I see two options:
> * cheaper: decrease the cores from 16 to 8 and push bonded calculations
> to gpu using gmx 2019
> * expensive: replace the 16core 2950X to 32core 2990WX
>
> 2950X 16 cores 2 GPUs
> gmx 2018 43 ns/day
> gmx 2019 43 ns/day
>
> 33 ns/day (8core/2GPUs) <<<< 54 (16core/4GPUS)
> 43 ns/day << 54 (16core/4GPUS)
>
> So this could be a compromise if 16/32 cores works similarly as 16/16
> cores. E.g. 2990 has slower memory access compared to 2950; I do not
> expect this to influence gmx runs too much. However, if it decreases by
> 10-15 percentage then most likely it does not worth to invest into the
> 32 core processor.
>
> Thanks for your feedbacks.
> Tamas
>
> --
> Tamas Hegedus, PhD
> Senior Research Fellow
> Department of Biophysics and Radiation Biology
> Semmelweis University     | phone: (36) 1-459 1500/60233
> Tuzolto utca 37-47        | mailto:tamas at hegelab.org
> Budapest, 1094, Hungary   | http://www.hegelab.org
> --
> 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.
>