lammps becom slowly

Dear all,
The lammps runs more and more slowly and uses more and more memory,what’s worse it apesrs that only one process running(one is in “D”,one is in “R”).
process.JPG
My in.file is:

Dear all,
The lammps runs more and more slowly and uses more and more memory,what's worse it apesrs that only one process running(one is in "D",one is in "R").

[...]

"mpirun -np 2 lmp_linux -sf cuda -var gpu_0 0 -var gpu_1 1 -in polymer.lmp"
It still output thermo data.

How to solve this problem?

hard to say. you are writing to the dump file quite frequently. have
you checked the size of the dump file and particularly if you have
free disk space?

axel.

I have enough free disk space.
I will write 80465 dump files ( 0.7M/file ) and I have already write 80177 files.

I have to write many dump files to calculate dynamic property.
Could you give me some advice?

Thank you very much.

I have enough free disk space.
I will write 80465 dump files ( 0.7M/file ) and I have already write 80177 files.

I have to write many dump files to calculate dynamic property.
Could you give me some advice?

not really. it is practically impossible to remotely determine what is
causing the issue. you need to find somebody that knows how to debug
and trace a running executable (there may be some tutorials on the
web, but i do not know).

axel.

Dear Dr. Axel Kohlmeyer

I convert the restart file to ordinary text file,then run my script without cuda package.The programs work very well and don’t request more and more memory.So I think it cause by cuda package.I don’t know how to debug cuda program,so I cannot pick up where it goes wrong.

I am using “lammps-12May13”,

I attach my in file and text file.

best wish


[polymer.lmp|attachment](upload://A0Oe5X7k4YPajugzHbEBPpRcTPk.lmp) (796 Bytes)



[restart.data|attachment](upload://9BzwHcTr07ZLAODyAAaivFl3k28.data) (1.74 MB)

Have you considered using the GPU package instead?

Also, before anybody will even consider looking into this, you should update tonthe current version of lammps and confirm that the issue is still reproducible.

Axel.

We use PADB
Padb: A parallel debugging tool
http://padb.pittman.org.uk/

It can give a snapshot of which part of the code each proc is in.

Nigel

Dear Dr. Axel Kohlmeyer

Using the latest version of LAMMPS (lammps-27Nov13) with cuda package (compiling with default configure) and the infile in bench/GPU/ directory(in.lj.cuda), LAMMPS uses more and more memory (writable memory and resident memory) ! More frequently the neighbor lists built,it need more memory!Memory leak emerge when neighbor list is built.But I can’t locate which function casue this problem.

Though it consume a little memory each time, LAMMPS will use up memory at last and become slowly!

Thank you!

please note that the USER-CUDA package is maintained by christian
trott (in cc) and even if i wanted to spend some time on this, i don't
have the time during the next week, since i am conducting an HPC
workshop in indonesia currently and will be on the road some more
after that.

have you tried using the GPU package instead?

axel.