[gmx-users] Fwd: Trouble with mpirun gmx- 2/2 machines not sharing correctly

[Mark Abraham]

Hi,

On Thu, Mar 10, 2016 at 7:24 AM Jacob Nowatzke <jn681 at humboldt.edu> wrote:

> I've recently installed OpenMPI 1.10.2 alongside GROMACS 5.1.2. I've read
> all over the web about how to, but I'm having issues left and right. I've
> finally gotten to a point where *gmx* works, but in an odd way. I'll try to
> explain without too much detail.
>
>
> Machine1=
> CPU: 8-core AMD FX9590
> GPU: AMD HD8990 (dual-core)
> RAM: 32GB
>
> Machine2=
> CPU: 6-core AMD FX6120
> GPU: AMD R9 Nano (single-core)
> RAM: 8GB
>
> Machines are networked by LAN, which cat I don't remember (may be reason
> for slow mdrun described soon?). I have added so many $PATH variables on
> and off- I'm tired of it. I do have to ask if I have to add
> usr/local/gromacs/bin to the $PATH and $LD_PATH in .bashrc? When I did this
> is when the machines started picking up on eachother, as they would not
> when only OMPI was specified in .bashrc. What's going on here?
>

I don't know, and probably nobody here knows how to set up and troubleshoot
a multi-node MPI setup, but you need to see a simple MPI program working
correctly across the two nodes before you even think about getting GROMACS
involved. Some observations that might lead you to avoid wasted effort are:

1) GROMACS won't benefit from multi-node runs unless your communication
environment has low latency, e.g. Infiniband, or possibly Gigabit ethernet,
so long as the network switch won't have lots of traffic on it (so probably
forget about it, particularly if it also supports your department's file
server)

2) GROMACS MPI runs between two nodes of different capability can work, but
getting best performance will require some juggling, and this is not a
great place for a beginner to start.

3) If you can justify doing two different simulations (and almost certainly
you can), one on each node, you don't need to do any of the above.


> My issue- besides understanding it all, this is the point where I've got
> things somewhat working- is that I run something like "mpirun --hosts
> Machine1,Machine2 --path usr/local/gromacs/bin gmx mdrun -v -deffnm nvt"
> from Machine2 over ssh from separate laptop, everything seems to be going
> smoothly... except this is only being run on Machine1(8990 is also
> detected), with Machine2 components not being detected. Essentially, I'm
> running gmx mdrun on Machine1 as if I was using ssh from Machine2.
> Obviously, this isn't what I want to do. I'd like to see each machine with
> their respective GPUs being utilized for gmx mdrun and the such.
>
> If Machine1 is actually being utilized, then why is it performing steps at
> a much slower rate than if I had used the same machine from ssh?
>

Either the network is too slow to be useful, or the setup isn't correct to
take advantage of it. Finding how how and why probably isn't useful, as
above.


> To top off this issue, when I run gmx grompp (or genion) it freezes before
> bringing me back to the command line.
>

Only mdrun might benefit from running across multiple ranks, so just don't
do that.

Mark

Obviously, I'm new, so I want to say that I really appreciate any help
> here. Learning GROMACS has been a pleasure so far and I'm hoping to get
> details down to utilize for homology modeling. Thanks for your help
>
> -Jacob
> --
> 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.
>