ARMCI error


Clicked A Few Times
Hi,

I am trying to run a DFT single point calculation with roughly 2000 electrons. The calculation stops before the first iteration with the following error :

Loading old vectors from job with title :



  Time after variat. SCF:    177.8
Time prior to 1st pass: 178.1

Grid_pts file          = ./1Auarmchair9x17.gridpts.00
Record size in doubles = 12289 No. of grid_pts per rec = 3070
Max. records in memory = 76 Max. recs in file = 53809113

0: error ival=4
(rank:0 hostname:node-e3-28 pid:596):ARMCI DASSERT fail. openib.c:armci_call_data_server():2010 cond:(pdscr->status==IBV_WC_SUCCESS)

I am running the calculation on 48 cpus with a total of 80 Go. I tried increasing the global memory to 60 Go without success.

Clicked A Few Times
I also added the following commands to my script

setenv ARMCI_DEFAULT_SHMMAX 2048
unsetenv MA_USE_ARMCI_MEM

What else could the problem be? Thanks in advance,

Chloe

Forum Vet
Please post the input file
Chloe,
Could you please post the complete input file (with basis set and -- if available -- ecp fields, of course)

Thanks, Edo

Clicked A Few Times
Here's a typical input file. First, I calculate the electronic structure with Hartree-Fock to ensure convergence. For that, I use a fragment guess. This calculation is fine. It's the DFT that crashes. I tried using an atomic guess in the DFT block without success. If you want the full input file with the geometries, I can send it by email (it is quite long).

echo
start
memory 80000 mb

geometry full units angstrom noautoz nocenter
...
symmetry c1
end

geometry frag1 units angstrom noautoz nocenter
...
symmetry c1
end

geometry frag2 units angstrom noautoz nocenter
...
symmetry c1
end

basis
 C library 3-21g*
H library 3-21g*
Au library lanl2dz_ecp
end

ecp
 Au library lanl2dz_ecp
end

set geometry full
set tolguess 1e-7
scf
direct
vectors fragment frag1.movecs frag2.movecs output HF.movecs
maxiter 400
end
task scf

dft
direct
xc b3lyp
vectors input HF.movecs output full.movecs
iterations 700
convergence ncyds 700
print "final vectors"
task dft energy

Forum Vet
Could you send the full input deck to bert.dejong@pnnl.gov and edoardo.apra@pnnl.gov so we can try to run it and see.

Bert


Quote:Chloe Feb 28th 8:04 pm
Here's a typical input file. First, I calculate the electronic structure with Hartree-Fock to ensure convergence. For that, I use a fragment guess. This calculation is fine. It's the DFT that crashes. I tried using an atomic guess in the DFT block without success. If you want the full input file with the geometries, I can send it by email (it is quite long).

echo
start
memory 80000 mb

geometry full units angstrom noautoz nocenter
...
symmetry c1
end

geometry frag1 units angstrom noautoz nocenter
...
symmetry c1
end

geometry frag2 units angstrom noautoz nocenter
...
symmetry c1
end

basis
 C library 3-21g*
H library 3-21g*
Au library lanl2dz_ecp
end

ecp
 Au library lanl2dz_ecp
end

set geometry full
set tolguess 1e-7
scf
direct
vectors fragment frag1.movecs frag2.movecs output HF.movecs
maxiter 400
end
task scf

dft
direct
xc b3lyp
vectors input HF.movecs output full.movecs
iterations 700
convergence ncyds 700
print "final vectors"
task dft energy

Forum Vet
Also, what platform are you running on?

Bert

Clicked A Few Times
It's an Infiniband cluster running Scientific Linux 6.1.

I realize this problem only occurs with large DFT calculations. Smaller ones end properly.

Chloe

Forum Vet
Please send us the large input deck that fails. How many processor cores were you running this on?

Bert

Quote:Chloe Mar 1st 2:00 pm
It's an Infiniband cluster running Scientific Linux 6.1.

I realize this problem only occurs with large DFT calculations. Smaller ones end properly.

Chloe


Forum >> NWChem's corner >> Running NWChem