[gmx-users] restart

[Yang Ye]

Okay. I think that optimize_fft=yes has caused the problem (indeed not a
problem). Setting optimize_fft=yes will call some benchmark function in FFTW
to test which "core" to use for FFT. The performances of these cores are
similar (I suspect) so this resulted subtle changes in the final choice on
some CPUs. P4 has this symptom, so derived Xeon may have, too.

A note to all these efforts in continuation. I know that such cautions shall
be a must. But for general-purpose protein folding, this seems not cause
much problem as I saw people publishing and got promising results without
exact continuation (just restart from a .gro with start-constraint=yes).
With increasing long-time simulation and extensive sampling, as well as
techniques like REME, this became less influential.

Nevertheless, I welcome to have it transparent to users (shuffle, sort,
parallel, etc...). Kudoz to the developers. Hope users are not spoiled with
the upcoming "holy grail" of MD.

Yang Ye

> -----Original Message-----
> From: gmx-users-bounces at gromacs.org [mailto:gmx-users-bounces at gromacs.org]
> On Behalf Of wnoid at hec.utah.edu
> Sent: Friday, February 10, 2006 11:38 PM
> To: gmx-users at gromacs.org
> Subject: [gmx-users] restart
> 
> hello all.
> 
> first of all, many thanks to mr. yang ye for his help with my earlier
> question.  (although, i am still curious to know exactly what grompp was
> doing.)
> 
> i have a question very closely related to yujie's and somewhat related to
> mr. mark abraham's comments.  i am trying to ensure the continuity of
> restarts of 4cpu bilayer simulations.  i have written a very simple script
> to check this on our local cluster of 64 bit amd opteron
> processors.  i compiled gromacs-3.3 in single precision on this cluster
> using gnu compilers and i find that continuing ten 1 step simulations with
> tpbconv agree with one 10 step simulation to within 1 part in 1e7.
> 
> however on a cluster of 32 bit intel xeon cpus my test fails very badly.
> on this cluster i compiled gromacs-3.3 in double precision using an intel
> compiler.  when i compare two 1 step simulations (continued with tpbconv)
> with one 2 step simulation i find that there are differences as large as 1
> part in 1e3 in the trr files, even though the starting tpr files for both
> sets of simulations are identical to within 1 part in 1e7.  from the trr
> files i find that after one step there are differences in the forces and
> after two steps differences show up even in the gro file output.  i am not
> using temperature or pressure coupling and am updating the neighbor list
> every step.
> 
> i start both simulations with grompp according to:
>       $grompp -f $gr_inp -c $config -p $topol\
>         -shuffle -np $nnodes\
>         -o $gr_out -po $gr_sum -deshuf $gr_des >& $gr_log
> 
> i extend the 1 step simulation using tpbconv according to:
>       $tpbconv -s $oldtpr -f $oldtrr -e $oldedr -extend $DeltaT\
>          -o $newtpr >& $tpblog
> 
> my grompp input file for the 1 step simulation is the following:
> (the input for the 2 step simulation is identical except nsteps =  2)
> 
> title               =  two 1step test
> cpp                 =  /lib/cpp
> constraints         =  all-bonds
> integrator          =  md
> dt                  =  0.001
> tinit               =  0.0
> nsteps              =  1
> nstcomm             =  1
> nstxout             =  1
> nstvout             =  1
> nstxtcout           = 1
> xtc_precision       = 1000
> nstfout             =  1
> nstlog              =  1
> nstenergy           =  1
> nstlist             =  1
> ns_type             =  grid
> coulombtype         = pme
> rlist               =  1.0
> rcoulomb            =  1.0
> rvdw                =  1.8
> fourierspacing      = 0.08
> pme_order           = 6
> ewald_rtol          = 1.E-5
> optimize_fft        = yes
> Tcoupl              =  no
> Pcoupl              =  no
> gen_temp            =  308.0
> gen_seed            =  48123
> pbc = xyz
> 
> 
> thanks to anyone who took the time to look at this.  any help would be
> greatly appreciated.
> 
> peace,
> 
> will noid
> 
> 
> 
> 
> 
> 
> 
> _______________________________________________
> gmx-users mailing list    gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/mailing_lists/users.php