2:19:50 PM PDT - Fri, May 18th 2012 |
|
mbpt4
|
Hi,
the mbpt2/mbpt3/mbpt4 implementations of NWChem are by-products of the CC iterations.
For this reason the mbpt2/mbpt3/mbpt4 implementations are not iptimal. In particular mbpt4
is memory thirsty, so be prepared to use large number of cores. Please use the following form of the
input:
tce
tilesize 8
mbpt4
end
(The mbpt-n implementations are supported by the genuine 4-index transformation only).
I made tilesize smaller because the local memory requirement of the 4-th order is proportional to
tilesize^8. I guess your problems with your job were associated with the excessive local memory requirements (for tilesize=10, tilesize^8 is equivalent to 0.75 GB).
You may want also to consider a cheaper form of the mbpt4 (the mbpt4(sdq) approach) which avoids
the storage of triply excited amplitudes in the second order. In this case please use the following input for the tce group
tce
tilesize 8
mbpt4(sdq)
end
Please also modify your memory settings:
memory stack 1500 mb heap 50 mb global 1400 mb
Best,
Karol
|