An advance warning: this will be an involved response.
On Sun, Apr 15, 2012 at 10:00 PM, 板凱哲彦 <itagaki08.gmail.com> wrote:
> Dear AMBER users and developers,
>
> I would like to ask partitioning issue of sander. I noticed that similar
>
> questions invoked in this ML forum, but could not find the solution.
>
> So, please let me allow to invoke this issue again.
>
>
> If you know the page providing the solution, please let me know just
>
> the page in this ML.
>
>
> I used sander to perform MD of protein using 8-core parallel computer
> system, and detected the following error message;
>
>
> ----------------------------------------
>
> partition error in shake on processor 0
> this processor has atoms 1 through 9100
> atom 9098 is within this range
> atom 9101 is not within this range !
>
> ------------------------------------------
>
This is a strange error, and one I've never seen with any of my systems.
To me, what this implies is a strange topology. Since I don't know any
details about how you created your topology file, I will simply provide as
much information about what causes this error as I can.
The base cause has to do with the way that sander is parallelized (i.e.,
how it distributes the workload across all processors). Sander has a
static load balancer (that is, it sets the workload at the beginning of the
simulation for each processor, and that never changes), and the workload is
split up by residue--each thread gets a number of whole residues that it is
principally responsible for. Therefore, I would fully expect this error to
depend on the number of threads you were running (since that changes how
the workload is distributed).
Because of the way SHAKE was programmed in sander, both the hydrogen and
the heavy atom involved in the constrained bond must be 'owned' by the same
thread (as described above). For 99.9% (99.999%?) of all residues, this is
no problem, since residues are almost always separated along heavy
atom-heavy atom bonds. The *only* way this error could appear is if a
residue division was defined along a heavy atom-hydrogen atom bond. This
would imply that the hydrogen atom is its own residue since H only forms 1
bond (let's pretend exceptions to this rule don't exist), or that the
topology file got messed up.
(Did you try modifying the topology file yourself?)
An alternate explanation has to do with ATOMS_PER_MOLECULE and
SOLVENT_POINTERS -- there are some corner cases where LEaP actually gets
these sections wrong. These sections only pertain to constant pressure
simulations, since only constant pressure makes use of 'molecule'
definitions (for the pressure scaling). Sander and pmemd have always just
assumed these sections would be correct (although I think Amber 12 checks
these sections now).
If you don't know which explanation is correct, try turning off constant
pressure (if constant pressure wasn't on, then it's the first explanation).
If the problem goes away after turning off constant pressure, I would
venture to say it's probably the first.
Also, try running the simulation with pmemd. If you still get problems
with pmemd (I've gotten segfaults using topologies with a bad
ATOMS_PER_MOLECULE section with pmemd), it's probably the second issue
(since pmemd has a completely different parallel scheme). If pmemd works,
it's probably the first issue.
If it's the second issue, there is a topology file 'doctor' of sorts
packaged with AmberTools 12 called ParmEd. Having noticed that LEaP
occasionally borks the ATOMS_PER_MOLECULE definition, I added a command to
fix topology files that have this issue. Instructions for running ParmEd
and the ParmEd GUI are in the AmberTools manual.
HTH,
Jason
P.S. If you are modifying topology files by hand, all bets are off, and
your problem could be anything. I highly suggest against modifying
topologies by hand -- try using ParmEd instead. It has many options for
customizing and modifying topology files in a way that should continue to
work with Amber programs.
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sun Apr 15 2012 - 21:00:03 PDT
