The ABINIT Users Mailing List ( CLOSED )

[Zeila Zanolli <zeila.zanolli@uclouvain.be>]

Hi Marc,

Thanks for the prompt answer: it clarify the difference between the serial and parallel 'estimates'.

I'm trying now the spin-polarized case without iprcel 45 and it is using ~ 2.7 Gb per cpu (12 cpus job).

Anyway, there is an imprecision in my previous email.

Concerning the spin-polarized case (with iprcell 45), I could check only now the actual use of memory per cpu:

in the parallel case, with 12 cpus, it is 10.6 Gb per cpu  (not 4.4 Gb as previously mentioned)

in the serial case it is still   10.6 Gb.

All the best,

Zeila

On 8 Jan 2010, at 18:02, TORRENT Marc wrote:

Hi Zeila,

I can partly answer to your question.

Don't take into account the account of memory announced by abinit in the "paral_kg=1" case.
It is probably completely wrong because not implemented for the parallel LOBPCG algorithm.


Other remark: iprcel 45 algorithm requires a huge amount of memory when called in band-fft parallel.
It could be an explanation..


Marc


Le 01/08/10 17:48, Zeila Zanolli a écrit :

C7C4FBAF-EA77-4332-A7C5-82D61B132B63@uclouvain.be" type="cite">

Dear all,

I'm trying to relax a ~60 atoms molecule with abinit and I would like to use the parallelization over the bands and fft to speed-up the calculations. 

However, it turned out that when the program is run in parallel mode it uses much more memory than in the serial mode!

I'll go for more details right now.

I'm running abinit-6.0.0 on a xeon cluster with ifort 10.1.

I've used the "paral_kgb = -n "  to chose the values of the npkpt, npband, npfft, bandpp for n cpus finding for 20 cpus, so the my input is:

paral_kgb 1 

npkpt      1

npband   1

npfft      20

bandpp   4

wfoptalg 4  nloalg 4  fftalg 401  intxc 0  fft_opt_lob 2

nband 100

kptopt 1

nshiftk 1

shiftk 0 0 0

ntime 1000

tolmxf 1.e-5

ionmov 2

ecut 35

nstep 100

iprcel 45

When running the code in serial mode, the estimation of the memory needed is

"This job should need less than    4.7 Gb of memory. "

and this corresponds what the code actually needs. (so this is ok).

When I'm running on 20 cpus the expected memory need is:

"This job should need less than    766.907 Mbytes of memory. "

but the code actually uses ~2.9  Gb per cpu (for a total of ~ 58 Gb of memory).

So.... why is it so? 

I also have another example, spin-polarized this time, where the situation is also bad.

For 12 cpus, the paral_kgb gives:

npkpt     2

npband  3

npfft       2

bandpp  2

In the serial mode the memory need is ~10.3 Gb  (expected memory need is 10.7 Gb in the output file).

When running on 12 cpus, the expected memory need (from the output file) is ~ 2.3 Gb

but the actual memory used is ~ 4.4 Gb per cpu, for a total of ~52.2 Gb 

(and then the code crashed since it overcomed the maximum memory of the machine).

Any help will be appreciated!

Thanks,

   Zeila

---------------------------------------------------------------------------------------------

Dr. Zeila Zanolli

Université Catholique de Louvain (UCL)

Unité Physico-Chimie et de Physique des Matériaux (PCPM) 

Place Croix du Sud, 1 (Boltzmann)

B-1348 Louvain-la-Neuve, Belgium

Phone: +32 (0)10 47 3501 

Mobile: +32 (0)487 556699

Fax: +32 (0)10 47 3452

e-mail: zeila.zanolli@uclouvain.be

web: http://www.nano2hybrids.net/

---------------------------------------------------------------------------------------------

---------------------------------------------------------------------------------------------

Dr. Zeila Zanolli

Université Catholique de Louvain (UCL)

Unité Physico-Chimie et de Physique des Matériaux (PCPM) 

Place Croix du Sud, 1 (Boltzmann)

B-1348 Louvain-la-Neuve, Belgium

Phone: +32 (0)10 47 3501 

Mobile: +32 (0)487 556699

Fax: +32 (0)10 47 3452

e-mail: zeila.zanolli@uclouvain.be

web: http://www.nano2hybrids.net/

---------------------------------------------------------------------------------------------