Molcas Forum

leo.c

Member

Registered: 2021-03-23

Posts: 5

Estimated memory requirements (CASPT2)

Hello everyone,

As I was trying to estimate the memory requirements for a future job (CASPT2), I failed to understand the relation between the given estimated memory requirements and the assigned value of MOLCAS_MEM (=112500 MB in the present case using only one thread on one process).

 Estimated memory requirements:
  POLY3 :              12590917656
  RHS:                     1324472
  SIGMA :                  2359744
  PRPCTL:                        0
 Available workspace:  12647109326

In which units are expressed these quantities ? I would have assumed bytes if nothing is specified but if so the value of the MOLCAS_MEM variable is about 10 times greater than the available workspace value given. I do not understand the difference here.

By the way I am using an openmp version with no GA of OpenMolcas 21.02.

Ignacio

Administrator

From: Uppsala

Registered: 2015-11-03

Posts: 1,011

Re: Estimated memory requirements (CASPT2)

I'd guess it's in 8-byte words (the space a single 64-bit integer/real takes)

leo.c

Member

Registered: 2021-03-23

Posts: 5

Re: Estimated memory requirements (CASPT2)

Indeed, order of magnitude checks out with 8-byte words. Thank you I should have checked that to begin with.

I have an additional question, perhaps a little off-topic. For such jobs, where multiple cores are usually not an option due to the high amount of memory needed, I wondered if there was a significant gain from using a serial build of OpenMolcas instead of the OpenMP build on 1 core. I have read something pointing in that direction in an old post, but I was curious to know if that was still relevant for the current version.

Ignacio

Administrator

From: Uppsala

Registered: 2015-11-03

Posts: 1,011

Re: Estimated memory requirements (CASPT2)

OpenMP parallelization (multiple threads) does not use more memory, at least not as far as MOLCAS_MEM is concerned, the memory is shared by all threads. It happens within a single process/node, and only (so far) in the linear algebra libraries (e.g. MKL). It should mostly be safe and faster than a single thread.

MPI/GA parallelization (multiple processes) runs a separate copy of the program on each process, which can be on different physical nodes. The MOLCAS_MEM value applies for each process, and each process uses its own separate memory.

Using a parallel (MPI) build of OpenMolcas with MOLCAS_NPROCS=1 should be equivalent to a serial build

leo.c

Member

Registered: 2021-03-23

Posts: 5

Re: Estimated memory requirements (CASPT2)

Thanks for clarifying these few points.