[lammps-users] failed to allocate %bytes...

LAMMPS LAMMPS Mailing List Mirror
1

Hi

My lammps simulations stopped suddenly with the error “failed to allocate/rellocate %bytes…” so that means I have to run it at higher number of processors or with fewer MPI tasks per node……it will change the trajectory as number of processors are changing……but will it affect the results such as diffusion or rdf?

Thanks

Chetan

2

Hi

chetan,

My lammps simulations stopped suddenly with the error "failed to
allocate/rellocate %bytes.." so that means I have to run it at higher

please quote error messages accurately (this is why 'cut-n-paste'
was invented in the first place), since the message you quote
does not exists in lammps and thus would originate most likely
from the MPI library and that in turn could have quite different
implications, e.g., depending on MPI implementation, lowlevel
communication hardware and settings, using more processors could
in some cases _increase_ the memory requirements.

it is important to understand why you run out of memory.
not using enough nodes to distribute the memory load is
only one of them and quite unlikely these days for classical
MD. to give you an example, in order to store one full set
of coordinates of a million particles in double precision
you need less than 25 Mbyte.

have you checked the memory usage of your MPI tasks?
how large are they? do they slowly grow over time?
what does the lammps memory usage statistics say?
could it be that you are running into a memory leak?

number of processors or with fewer MPI tasks per node..it will change the
trajectory as number of processors are changing..but will it affect the
results such as diffusion or rdf?

this question has been answered many, many times on this mailing
list before, on occasion quite detailed. so i suggest you first
search in the mailing list archives. regardless of that, this is
a question that you should be able to answer _yourself_ before
actually doing MD production runs.

cheers,
   axel.

3

In general, you should get statistically similar answers
running on different numbers of procs, but likely not
the exact same answers.

Steve