problem with "dump_modify append" when using MPI

LAMMPS LAMMPS Mailing List Mirror
1

Hello, Lammps users,

I want dump snapshots to be appended to the end of an existing dump file. So I am using the commond:

dump_modify 1 append yes.

But there is a problem with it, when I try to run the simulation on a parallel machine by MPI, the dump output file is in a mess, looks like:

ITEM: TIMESTEP
0
ITEM: NUMBER OF ATOMS
28314
ITEM: BOX BOUNDS pp pp pp
-14.5 14.5
-14.5 14.5
-14.5 14.5
ITEM: ATOMS id mol type xu yu zu vx vy vz
1 1 1 -13.9423 -13.9423 -9.2 1.93677 -0.508665 0.243565
2 1 1 -13.9423 -13.9423 -9.8 1.50259 -0.535623 0.243565
3 1 1 -13.9423 -13.9423 -10.4 1.06841 -0.562582 0.243565

......

606 68 1 2.78846 -11.7115 -10.4 -0.591271 -0.80807 0.603365
60ITEM: TIMESTEP
0
ITEM: NUMBER OF ATOMS
28314
ITEM: BOX BOUNDS pp pp pp
-14.5 14.5
-14.5 14.5
-14.5 14.5
ITEM: ATOMS id mol type xu yu zu vx vy vz
1 1 1 -13.9423 -13.9423 -9.2 1.93677 -0.508665 0.243565
2 1 1 -13.9423 -13.9423 -9.8 1.50259 -0.535623 0.243565
3 1 1 -13.9423 -13.9423 -10.4 1.06841 -0.562582 0.243565

......

I guess all porcessors participate in dump in a random order, and that makes a mess.

After I delete the command "dump_modify append" in my script, everything is good again.

Thank you very much,

Best wishes,

Wei Chen

2

Hello, Lammps users,

I want dump snapshots to be appended to the end of an existing dump file. So I am using the commond:

dump_modify 1 append yes.

this is almost always a bad idea.
it is better to have each run write to a
different file in order to avoid problems
with file corruption on networked file
systems. dump files can always be
concatenated at a later stage.

But there is a problem with it, when I try to run the simulation on a parallel machine by MPI, the dump output file is in a mess, looks like:

please check the lammps output carefully.
this looks almost like you are trying to run
a serial executable in parallel. also you have
to make sure that any previous run was not
terminated prematurely and left you with an
incomplete file.

ITEM: TIMESTEP
0
ITEM: NUMBER OF ATOMS
28314
ITEM: BOX BOUNDS pp pp pp
-14.5 14.5
-14.5 14.5
-14.5 14.5
ITEM: ATOMS id mol type xu yu zu vx vy vz
1 1 1 -13.9423 -13.9423 -9.2 1.93677 -0.508665 0.243565
2 1 1 -13.9423 -13.9423 -9.8 1.50259 -0.535623 0.243565
3 1 1 -13.9423 -13.9423 -10.4 1.06841 -0.562582 0.243565

......

606 68 1 2.78846 -11.7115 -10.4 -0.591271 -0.80807 0.603365
60ITEM: TIMESTEP
0
ITEM: NUMBER OF ATOMS
28314
ITEM: BOX BOUNDS pp pp pp
-14.5 14.5
-14.5 14.5
-14.5 14.5
ITEM: ATOMS id mol type xu yu zu vx vy vz
1 1 1 -13.9423 -13.9423 -9.2 1.93677 -0.508665 0.243565
2 1 1 -13.9423 -13.9423 -9.8 1.50259 -0.535623 0.243565
3 1 1 -13.9423 -13.9423 -10.4 1.06841 -0.562582 0.243565

......

I guess all porcessors participate in dump in a random order, and that makes a mess.

your guess is wrong. in this style of dump, output is done
by only one processor.

After I delete the command "dump_modify append" in my script, everything is good again.

can you demonstrate this by modifying one of the lammps examples,
e.g. the melt example?

thanks,
    axel.

3

Hi, Axel,

Thanks for your reply. I tried the example melt with and without "dump_modify append"

the command is

dump_modify id append yes

I am sure all previous runs are finished. The dump file with "append" is still in mess,

ITEM: TIMESTEP
0
ITEM: NUMBER OF ATOMS
4000
ITEM: BOX BOUNDS pp pp pp
0 16.796
0 16.796
0 16.796
ITEM: ATOMS id type xs ys zs
1 1 0 0 0
2 1 0.05 0.05 0

......

1501 1 0.5 0ITEM: TIMESTEP
0
ITEM: NUMBER OF ATOMS
4000
ITEM: BOX BOUNDS pp pp pp
0 16.796
0 16.796
0 16.796
ITEM: ATOMS id type xs ys zs
1 1 0 0 0
2 1 0.05 0.05 0

......

Best wishes,

Wei

4

The dump file with "append" is still in mess,

What does this mean?

I just ran bench/in.lj and added these lines:

dump 1 all atom 10 tmp.dump
dump_modify 1 append yes

I get the same dump file whether I comment out
the 2nd line or not. It works in serial or parallel.

Steve

5
6

and would you please post the log file that you create when running in parallel.
i maintain, that what you are showing is most likely to happen with an
incorrectly compiled executable. i could reproduce this on my local machine
when running a serial executable with 8 copies.

thanks,
     axel.

7

i send you a sent of file in a private e-mail.
no need to spam the list with this.

axel.

8

Hi, Axel,

Thank you for your reply.

You are right, I checked the log file in parallel with 4 processors, but it is said

1 by 1 by 1 MPI processor grid

......

Loop time of 1.08555 on 1 procs for 250 steps with 4000 atoms

I am running a serial executable by mistake.

Thanks again,

Best wishes,

Wei

9

Hi, Axel,

It is clear now, there are two different MPI in workstation, when lammps is built, it finds the files of MPI A, but when I submit the job, system uses MPI B to run lammps. so it is regarded as a serial program by MPI B.

Best wishes,

Wei