ga scatter acc:invalid i/j Error

Just Got Here

4:18:14 AM PDT - Thu, May 15th 2014

Hi,

I'm currently doing some CASSCF calculation on CH+ system with Nwchem.
With this input:

title "ch+_0.04"
scratch_dir /home/ftalotta/scratch/
memory heap 100 mb stack 1200 mb global 600 mb
geometry units an print
 c 0.0000	 0.0000	 0.0000
 h 0.64	 0.0000	 0.0000
 symmetry c1
end
 
basis
 c library aug-cc-pvtz
 h library aug-cc-pvtz
end

charge 1

set lindep:tol 1.0d-7

scf
 singlet
 rhf
 maxiter 1000
end
 
task scf energy

mcscf
 state 1a
 active 15
 actelec 4
 maxiter 100
end
 
task mcscf energy

and I get this error:

1:ga_scatter_acc:invalid i/j=(93,-1):Received an Error in Communication
--------------------------------------------------------------------------
MPI_ABORT was invoked on rank 1 in communicator MPI COMMUNICATOR 3 DUP FROM 0 
with errorcode -993.

It seems there is something wrong in the ga_scatter_acc routine.

How to fix that?

Thanks

Francesco

Forum Regular

9:40:40 PM PDT - Thu, May 15th 2014
Hi Francesco, How many cores are you running this calculation on? I have just run this calculation on my desktop machine using 3 cores and that worked fine. I am wondering if you are using so many cores that some don't hold any data. It is possible that some Global Array accesses have not been protected with that possibility in mind. Nevertheless, if you can tell how many cores you are using I can see where it goes wrong and fix that. Huub

Just Got Here

3:16:42 AM PDT - Fri, May 16th 2014
Hello, thanks for reply. I did it with 4 cores, but now I tested it with 3 cores, with the same results! Francesco Quote:Huub May 16th 4:40 am Hi Francesco, How many cores are you running this calculation on? I have just run this calculation on my desktop machine using 3 cores and that worked fine. I am wondering if you are using so many cores that some don't hold any data. It is possible that some Global Array accesses have not been protected with that possibility in mind. Nevertheless, if you can tell how many cores you are using I can see where it goes wrong and fix that. Huub

Forum Regular

5:00:46 PM PDT - Tue, May 20th 2014
Hi Francesco, I am having a hard time trying to reproduce the error that you are seeing. With the input you provided i/j in ga_scatter_acc should never exceed 10. Yet in your case the code complains that pair i/j = 93/-1 are out of range (and that is correct but I don't get where that i/j combination is coming from). Can you send me the full output, please? Send it to hubertus dot vandam at pnnl dot gov. Maybe there is a hint of what is going wrong some where in there. Huub

Forum Regular

1:51:32 PM PDT - Wed, May 21st 2014

Hi Francesco,

The problem is that you are using 15 active orbitals whereas the code, as distributed, only allows for 14. In the development version this problem is detected already, hence I was having trouble reproducing it. The NWChem-6.3 release code does not check whether the 14 active orbitals limit is exceeded. Instead the code suffers severe memory overwrites and (hopefully) fails badly at some point.

The fix is to edit

   nwchem-6.3/src/mcscf/detci/detciP.fh

and change the line that reads

   parameter(detci_maxorb=14)

Replace the number 14 with something a little larger (maybe as large as 30 as long as you are not going to try and use that many with more than a few electrons). Then run

   pushd $NWCHEM_TOP/src/mcscf/detci; make FC=$FC; pushd $NWCHEM_TOP/src; make FC=$FC link; popd; popd

and you should be good to go.

Let me know if you need any further assistance, please.

Best wishes, Huub

Just Got Here

2:05:30 AM PDT - Fri, May 23rd 2014
Hi Huub, I'll check that in a couple of days!! Many Thanks for your support. Francesco

Forum >> NWChem's corner >> Running NWChem