Re: AMBER: amber/pmemd with mpiexec

From: Robert Duke <rduke.email.unc.edu>
Date: Mon, 12 Apr 2004 10:41:22 -0400

Lubos -
I tend to defer to the core amber team on such issues. I don't see a big
problem with it; if your code works for you, that is all the approval you
need of the source below. I am not sure of the real need for such things,
and generally associate mixed parameters on the command line with bad script
construction, though if the scripts are part of a released system like pbs,
they could be hard to deal with. Ideally though, sander or pmemd or
whatever would be exec'd with a clean command line, and all the junk
associated with setting up the mpi environment would get filtered out. I
have not studied this extensively; perhaps someone on the amber team with a
history of fighting this stuff will have more cogent comments.
Regards - Bob

----- Original Message -----
From: "Lubos Vrbka" <shnek.tiscali.cz>
To: <amber.scripps.edu>
Sent: Sunday, April 11, 2004 10:00 AM
Subject: Re: AMBER: amber/pmemd with mpiexec


> bob,
>
> > If you really need this, for pmemd just edit get_cmdline.f90 to
> > include -execer_id as something that is filtered out in the mpi code.
First
> > grep for "-np", see what is done to filter it out, and do the same thing
> > for -execer_id. You can do the same thing in sander; the relevant code
is
> > in mdfil.f.
> thanks for hint... it is working now. since i don't know fortran at all,
> i'm not sure whether my modification is correct in the sense of syntax
> and fortran-programming habits, but anyway i post it here just in case
> it would be useful for someone in the future...
>
> this change is necessary (i think) only for openpbs/pbspro queuing
> system and submission with p4 mode of mpich via mpiexec program. shmem
> mode submission with mpiexec works fine.
>
> mpiexec doesn't use mpirun, but connects to appropriate nodes and start
> jobs by itself. it has some advantages over mpirun:
>
> - gets the information for all running processes directly from PBS, so
> it always has the correct information about e.g. cpu load, ...
> - it properly kills all jobs when requested by user. mpirun-started p4
> jobs that are killed by user often don't quit and remain as zombies on
> the computer nodes, ...
>
> ! modification to enable mpiexec submission/execution
> ! add after line 164 in the original pmemd-3.1 file get_cmdline.f90
> else if (arg .eq. '-execer_id') then
> iarg = iarg + 1
> else if (arg .eq. '-master_host') then
> iarg = iarg + 1
> else if (arg .eq. '-my_hostname') then
> iarg = iarg + 1
> else if (arg .eq. '-my_nodenum') then
> iarg = iarg + 1
> else if (arg .eq. '-my_numprocs') then
> iarg = iarg + 1
> else if (arg .eq. '-total_numnodes') then
> iarg = iarg + 1
> else if (arg .eq. '-master_port') then
> iarg = iarg + 1
> else if (arg .eq. '-remote_info') then
> ! format: -remote_info <hostname> <procs-on-that-node> ...
> iarg = iarg + 1
> ! to filter out all "hostname - procs" pairs
> ! we do it one by one till the '-' is the first
> ! character in the argument or end-of-parameters
> ! occurs
> do
> if (iarg .eq. indx) then
> exit
> end if
> call getarg(iarg,arg)
> if (arg(1:1) .eq. '-') then
> exit
> else
> iarg = iarg + 1
> end if
> end do
>
> is there any chance to have this supported in some future releases of
amber?
>
> regards,
>
> --
> Lubos
> _._"
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>


-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Mon Apr 12 2004 - 15:53:00 PDT
Custom Search