Memory requirements of EOM-CCSD excitations

From NWChem

Viewed 2128 times, With a total of 1 Posts
Jump to: navigation, search

Clicked A Few Times
Threads 6
Posts 13
Hello All,

I just started playing around with EOM-CCSD for excitation energies with NWChem 6.5. I have a smallish test system that just won't finish, apparently due to out-of-memory. Here's the input file I'm trying to run:

START phpyr
TITLE "PhPyr EOM-CCSD"

memory stack 3000 mb heap 600 mb global 3400 mb

GEOMETRY
C          0.00000        0.00000       -3.20391
C          1.20789        0.00000       -2.49014
C          1.21375        0.00000       -1.08902
C          0.00000        0.00000       -0.36623
C         -1.21375        0.00000       -1.08902
C         -1.20789        0.00000       -2.49014
N          0.00000        0.00000        1.05365
C          1.12627        0.00000        1.87026
C          0.71524        0.00000        3.19179
C         -0.71524        0.00000        3.19179
C         -1.12627        0.00000        1.87026
H          0.00000        0.00000       -4.30386
H          2.16819        0.00000       -3.02830
H          2.17635        0.00000       -0.56155
H         -2.17635        0.00000       -0.56155
H         -2.16819        0.00000       -3.02830
H          2.13532        0.00000        1.45205
H          1.37744        0.00000        4.06507
H         -1.37744        0.00000        4.06507
H         -2.13532        0.00000        1.45205
END

BASIS spherical
 * library def2-SVP
END

SCF
 SINGLET
 RHF
 thresh 1.0e-10
 tol2e 1.0e-10
END

TCE
 CCSD
 2eorb
 eomsol 2
 diis 5
 tilesize 10
 attilesize 30
 DIPOLE
 FREEZE CORE ATOMIC
 NROOTS 3
 TARGETSYM a1
 SYMMETRY  a1
END
set tce:thresheom 1.0d-4
set tce:threshl 1.0d-3


TASK TCE ENERGY


I have played around with different settings for "tilesize" and the partitioning of the memory, without success. The calculation always crashes at some point. With the above, it goes the furthest, starting to solve for "EOM-CCSD left-hand side iterations", crashing after iteration 13 with:

 Iteration  12 using   36 trial vectors
   0.0003636340752   0.1986052891396    5.40433
   0.0007455083161   0.2456298542250    6.68393
   0.0019875161423   0.2545313238981    6.92615    45.9    46.7

 Iteration  13 using   39 trial vectors
Application 2467260 exit codes: 134
Application 2467260 exit signals: Killed
Application 2467260 resources: utime ~2320s, stime ~49s, Rss ~38212, inblocks ~3087249, outblocks ~9166261


In stdout I get errors like:
 PE 284 [Wed Feb 25 04:07:16 2015] [c6-0c1s3n0] [nid01228] LIBDMAPP ERROR: Unsuccessful memory registration.
nwchem: ../../ga-5-3/comex/src-dmapp/comex.c:890: _comex_malloc_local: Assertion `status == DMAPP_RC_SUCCESS' failed.


If someone who knows the code and module much better than me could confirm that the calcs really are supposed to need this much memory, I can stop trying to play around, because then any "real" systems will be way beyond current resources. Even better, any suggestions on how to get CCSD excitation energies with less memory would be great!

The hardware I have access to is a Cray XC40, and I can run on 42 nodes with 24 cores and 64 GB of memory per node. The above example was run on 42 nodes utilising only 8 cores per node in order to have 8 GB/core, but that is of course not optimal. I'd rather run these types of calcs on 1008 cores with 2.6 GB/core, if at all possible.

Any and all help appreciated.

Clicked A Few Times
Threads 6
Posts 13
Just a short follow-up: Seems to be a GA-specific problem. When compiled with ARMCI/MPI or pure MPI, the calc goes through with less than 2GB/core. Slower, but finishes at least.


Forum >> NWChem's corner >> Running NWChem



Who's here now Members 0 Guests 1 Bots/Crawler 0


AWC's: 2.5.10 MediaWiki - Stand Alone Forum Extension
Forum theme style by: AWC