[gmx-users] mpi run in Gromacs 4.5.3

[Mark Abraham]

On 14/12/2010 11:14 AM, Te, Jerez A., Ph.D. wrote:
> Hi Mark,
>
> Thank you for your reply. Just to confirm, mdrun_mpi is still being used in Gromacs 4.5.3? The gromacs manual (Appendix A.5- Running Gromacs in parallel) suggested using "mdrun -np 8 -s topol -v -N 8" in running a single machine with multiple (8) processors and "mpirun -p goofus,doofus,fred 10 mdrun -s topol -v -N 30" to run three machines with ten processors each. You probably already know this but I am confused as to whether these commands are correct in running parallel simulations (thus the assumption that mdrun_mpi is no longer applicable in gromacs 4.5.3). So far I have not been successful in running parallel simulations (as I mentioned before the two options above would only start X identical serial processes). My bottom line is I want to run Gromacs simulation in parallel (regardless of whether it is running on one silicon with a number of processors or on different machines or nodes with a specified number of processors).
>
> I assume that in order to get mdrun_mpi, Gromacs has to be recompiled using the --enable-mpi option because currently our gromacs executable bin does not have mdrun_mpi. Our system administrator said that all mpi-related options have been enabled in compiling gromacs and still the mdrun_mpi is not found as one of the exe files. Please help.
I echo Justin's comments.

"all mpi-related options" seems suspicious. There's only one, 
--enable-mpi. Obviously you need a correctly-configured MPI compiler and 
environment for it to work

Mark
> Thanks,
> JT
>
>
> -----Original Message-----
> From: gmx-users-bounces at gromacs.org on behalf of Mark Abraham
> Sent: Mon 12/13/2010 4:38 PM
> To: Discussion list for GROMACS users
> Subject: Re: [gmx-users] mpi run in Gromacs 4.5.3
>
> On 14/12/2010 7:48 AM, Te, Jerez A., Ph.D. wrote:
>> Hi,
>> I have been trying to run Gromacs 4.5.3 parallel simulations using
>> openmpi 1.4.2. From my understanding, mdrun_mpi is not used in this
>> version of Gromacs.
>>
> I don't understand what (you think) you mean. You can use thread-based
> parallelism for processors that share common silicon, or MPI-based
> parallelism if a network connection is involved, but not both. The
> latter is named mdrun_mpi by default.
>
>> Our system administrator told me that all mpi related options have
>> been turned on while installing Gromacs. With either commands:
>> mdrun -np X -deffnm topol -N X (run in an 8-cpu node)
>>
> This won't run in parallel at all. mdrun ignores -np and -N
>
>> or
>> mpirun -np X mdrun -deffnm topol -N X (submitted in a number of nodes
>> depending on availability)
>>
> This will get you the symptoms below, but -N is still ignored.
>
>> I get X identical simulations instead of a parallel run. If X=4, I get
>> 4 identical simulations (the same simulation ran 4 times) instead of 1
>> parallel simulation in 4 processors. The performance difference
>> between a single-processor run and the X=4 run are also similar (no
>> marked difference in the time it takes to finish the simulation).
>> Has anyone encountered this problem?
>>
> You're using a serial mdrun. Use a parallel mdrun.
>
> Mark
>