Molcas Forum

tvery

Member

Registered: 2019-07-11

Posts: 6

Estimation of the memory required by a module

Hello,

I am trying to run a CASSCF computation with Molcas 8.4.

The job crashes because of memory limits.
   Maximal available memory for Molcas = 8879921020
   MEMORY ERROR: Memory is exhausted!
   MEMORY ERROR: Available memory = 8879921020 ( 8879 Mb ) !
   MEMORY ERROR: Requested memory = 12018733056 ( 12018 Mb ) !
   MEMORY ERROR: The suggested MOLCAS_MEM=53139 !
   C_GetMem Calling parameters: ('KICTS  ','ALLO   ','INTE   ',0,1502341632)

I tried to allocate up to 120GB/process.

Is there a way of estimating the memory needs of a module (rasscf in my case)?

Best regards

Thibaut Véry

Ignacio

Administrator

From: Uppsala

Registered: 2015-11-03

Posts: 1,085

Re: Estimation of the memory required by a module

Not that I know...

With 120GB you should not be getting that error. What does the module header say?

tvery

Member

Registered: 2019-07-11

Posts: 6

Re: Estimation of the memory required by a module

The header is the following.
I made a mistake I have 110 GB not 120.

()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()

                                              &RASSCF

             launched 32 MPI processes, running in PARALLEL mode (work-sharing enabled)
                       available to each process: 110 GB of memory, 1 thread?
()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()

The computation ran on 32 nodes with 1 MPI task on each.
Each node has 32 cores.

nikolay

Member

From: Stuttgart

Registered: 2016-03-21

Posts: 54

Re: Estimation of the memory required by a module

Could it be that something went wrong and 32 processes land on one node?
From my experience there is no parallel speedup for RASSCF (correct me if I'm wrong).
You could try to run this job in serial, this is also a test if 110 GB is enough.

tvery

Member

Registered: 2019-07-11

Posts: 6

Re: Estimation of the memory required by a module

@nikolay I think that the processes use the 32 nodes (from what I can see)

I can try to run the job serially.
Will it be OK with the OrdInt file?

I looked inside the temporary files generated and there is this message :

  *************************************************
  Sorry, but your calculation will probably be too
  large for the available memory.
  The number of determinants is             1502341632
  During CI equation solution, there will be
  up to six vectors of this size in memory.
  We estimate an additional 50% for other stuff.

 This alone will need at least  103214.3 MB.
 Available at this point is   48831.2 MB.
  Please increase MOLCAS_MEM, and try again.
  *************************************************

nikolay

Member

From: Stuttgart

Registered: 2016-03-21

Posts: 54

Re: Estimation of the memory required by a module

I think it should be fine to reuse the OrdInt file.
What is your active space?

tvery

Member

Registered: 2019-07-11

Posts: 6

Re: Estimation of the memory required by a module

Here is an excerpt from the input file.
-------------

Spin
 2
Symmetry
 1
Inactive
 222 195
Ras2
 11 9
*
Nactel
 13 0 0
*

----------

What do you mean reuse the OrdInt file?
I got one (indeed 32 since I use 32 MPI tasks) from seward.

nikolay

Member

From: Stuttgart

Registered: 2016-03-21

Posts: 54

Re: Estimation of the memory required by a module

With (13, 20) active space it should work fine with serial execution.

Will it be OK with the OrdInt file?

I thought you were reusing integral file to speedup the calculation.
If you don't do so, better use Cholesky decomposition to speedup integrals calculation by adding

RICD

in gateway.