Ray, hi.
Sorry, I forgot to copy my previous message to the mailing list. I'm fixing it now.
So, it works for around 4-5 thousand atoms and 1.3 GB memory per core? It is sufficient for actual calculations, I think, and it may be used.
I reproduced the problem with 5*10*10 cells on 4-core (8-thread) i7 machine with 12 GB memory. When I run 4 MPI processes, it gives seg fault at the very end of the calculation, after plotting log to the screen. Position files in this case are also incorrect. But on 5, 6, 7, 8 threads it produces a lot of "free(): invalid pointer" errors at the first output step (step 1). Don't know how to explain this yes, maybe it is hardware-dependent or library-dependent.
The version nlocal_max*MAX_NEIGH_ATOM doesn't work here as well. But it works when I replaced nlocal_max back to nmax, so it is nmax*MAX_NEIGH_ATOM. I suppose it is not the solution and there may be errors at larger scales, but such procedure worked for 2*2*2 system for 1000 timesteps. I attach this version to the message.
Oleg.
05.06.2013, 06:30, "Ray Shan" <[email protected]...>:
reaxc_species_test.tar.gz (7.92 KB)