large RIDFT calculations

[sfomine]

Dear All

I am still a newbie to TM facing a problem trying to run the calculations for a large molecular complex. I managed to run dscf, but ridft constantly crushes without any warning when reading initial guess orbitals. I tried both, normal ridft  and ridft_huge. Both failed to run. I even tried to use MOs from dscf run, nothing helps. The number of atoms of the complex is under 700 atoms. When I decrease the size of the complex the calculations run OK. $Ricore apparently does not accept more than 2000 mb memory, is it 32 bit application? I am attaching control and  slave1.output of the crushed job. I can not figure out what is wrong.  I use 64bt TM.6.4 on a CENTOS 6.4 x64 box. Thank you in advance. Any suggestions will be greatly appreciated.

Sincerely

Serguei  Fomine

National University of Mexico

Institute for material researches

[antti_karttunen]

Hi Serguei,

Ridft should work with this size of complex and also with more than 2GB of $ricore, so probably something is wrong with the parallel/memory settings. I noticed that you are running the Global array (GA) binaries. You could first try using the "normal" MPI binaries to check if the problem is related to the GA binaries. Another option would be to play with the GA memory settings (http://www.cosmologic.de/data/DOK_HTML/node29.html). From your slave1.output:
Default GA heap smaller than $ricore!
GA heap raised to: 311296000
 par_setmem:              311296000               1310720 F
 MA:  Allocated stack:    10.00000      MB/processor
 SHM: Allocated heap :    2375.000      MB/processor
Now, my guess is that the allocated stack space is way too low for > 6000 AOs. You could try increasing it to at least 100MB (or even 1000 MB?) per CPU (I'm not that familiar with the memory requirements of the GA version). If the memory requirements are too large for your hardware, just set $ricore to 0, this shouldn't slow the calculation too much. Finally, it might also help in debugging if you could get the possible fortran/system error messages from stderr/stdout.

Best,
Antti

[sfomine]

Hello Antti

Thank you very much. You were absolutely right the problem was a stacksize. I am still unable to run the task on 16 cores, but the task runs on 15 when the stacksize is increased to 5 GB.

Best regards

Serguei