[molpro-user] memory settings in parallel runs
Manuel Dömer
manuel.doemer at epfl.ch
Wed Nov 11 13:47:27 GMT 2009
Dear Molpro Users,
I am running Molpro (2006.1) CCSD(T) calculations with large basis
sets on an IBM P5 cluster with the LoadLeveler Batch System and I have
run out of memory with my settings. After reading the sensible parts
in both the Molpro and the LL description I am still a bit confused
and would kindly ask for some advice.
1. the dynamical allocatable memory per task is given by the -m
option. When the memory of one node is not enough I extend the number of
nodes, but then I also need to specify the global array memory
with the -G option. How can I find out whether I need more dynamical
memory or more global memory?
2. is there a way just to check the memory needs of a specific Molpro
job?
3. when I run out of memory, is it a good idea (in terms of the memory
problem, not the fact that I am wasting cpu time) to increase the
amount of dynamical memory per task while reducing the number of
tasks per node? Of course the total amount of dynamical memory stays
the same in this way.
4. are the following warnings and error messages related to the
dynamical memory (-m option) or to the global array memory (-G option)?
Molecular orbitals read from record 2100.2 Type=RHF/CANONICAL
(state 1.1)
-10000: WARNING:armci_set_mem_offset: offset changed 0 to 29634560
-10016: WARNING:armci_set_mem_offset: offset changed 0 to 29634560
-10032: WARNING:armci_set_mem_offset: offset changed 0 to 29634560
3-fold multipass in aabb transformation; to avoid, increase memory
by238434773
NPASS IN TRSOR3 3
-------------
::: For minimal I/O algorithm in triples, increase memory by
356884442 words to 592.2 Mword
::: For full I/O caching in triples, increase memory by
1266559116 words to 1501.8 Mword
::: A further 142942592 words of memory are needed for the
triples to run
::: Increase memory to 378.2 Mword
GA ERROR fehler on processor 0
I thank you very much in advance!
Kind Regards,
Manuel
More information about the Molpro-user
mailing list